This post was originally published on the author’s personal blog.

Last year’s Conference on Robot Learning (CoRL) was the biggest CoRL yet, with over 900 attendees, 11 workshops, and almost 200 accepted papers. While there were a lot of cool new ideas (see this great set of notes for an overview of technical content), one particular debate seemed to be front and center: Is training a large neural network on a very large dataset a feasible way to solve robotics?1

Of course, some version of this question has been on researchers’ minds for a few years now. However, in the aftermath of the unprecedented success of ChatGPT and other large-scale “foundation models” on tasks that were thought to be unsolvable just a few years ago, the question was especially topical at this year’s CoRL. Developing a general-purpose robot, one that can competently and robustly execute a wide variety of tasks of interest in any home or office environment that humans can, has been perhaps the holy grail of robotics since the inception of the field. And given the recent progress of foundation models, it seems possible that scaling existing network architectures by training them on very large datasets might actually be the key to that grail.

Given how timely and significant this debate seems to be, I thought it might be useful to write a post centered around it. My main goal here is to try to present the different sides of the argument as I heard them, without bias towards any side. Almost all the content is taken directly from talks I attended or conversations I had with fellow attendees. My hope is that this serves to deepen people’s understanding around the debate, and maybe even inspire future research ideas and directions.

I want to start by presenting the main arguments I heard in favor of scaling as a solution to robotics.

• It worked for Computer Vision (CV) and Natural Language Processing (NLP), so why not robotics? This was perhaps the most common argument I heard, and the one that seemed to excite most people given recent models like GPT4-V and SAM. The point here is that training a large model on an extremely large corpus of data has recently led to astounding progress on problems thought to be intractable just 3-4 years ago. Moreover, doing this has led to a number of emergent capabilities, where trained models are able to perform well at a number of tasks they weren’t explicitly trained for. Importantly, the fundamental method here of training a large model on a very large amount of data is general and not somehow unique to CV or NLP. Thus, there seems to be no reason why we shouldn’t observe the same incredible performance on robotics tasks.
  • We’re already starting to see some evidence that this might work well: Chelsea Finn, Vincent Vanhoucke, and several others pointed to the recent RT-X and RT-2 papers from Google DeepMind as evidence that training a single model on large amounts of robotics data yields promising generalization capabilities. Russ Tedrake of Toyota Research Institute (TRI) and MIT pointed to the recent Diffusion Policies paper as showing a similar surprising capability. Sergey Levine of UC Berkeley highlighted recent efforts and successes from his group in building and deploying a robot-agnostic foundation model for navigation. All of these works are somewhat preliminary in that they train a relatively small model with a paltry amount of data compared to something like GPT4-V, but they certainly do seem to point to the fact that scaling up these models and datasets could yield impressive results in robotics.
• Progress in data, compute, and foundation models are waves that we should ride: This argument is closely related to the above one, but distinct enough that I think it deserves to be discussed separately. The main idea here comes from Rich Sutton’s influential essay: The history of AI research has shown that relatively simple algorithms that scale well with data always outperform more complex/clever algorithms that do not. A nice analogy from Karol Hausman’s early career keynote is that improvements to data and compute are like a wave that is bound to happen given the progress and adoption of technology. Whether we like it or not, there will be more data and better compute. As AI researchers, we can either choose to ride this wave, or we can ignore it. Riding this wave means recognizing all the progress that’s happened because of large data and large models, and then developing algorithms, tools, datasets, etc. to take advantage of this progress. It also means leveraging large pre-trained models from vision and language that currently exist or will exist for robotics tasks.
• Robotics tasks of interest lie on a relatively simple manifold, and training a large model will help us find it: This was something rather interesting that Russ Tedrake pointed out during a debate in the workshop on robustly deploying learning-based solutions. The manifold hypothesis as applied to robotics roughly states that, while the space of possible tasks we could conceive of having a robot do is impossibly large and complex, the tasks that actually occur practically in our world lie on some much lower-dimensional and simpler manifold of this space. By training a single model on large amounts of data, we might be able to discover this manifold. If we believe that such a manifold exists for robotics — which certainly seems intuitive — then this line of thinking would suggest that robotics is not somehow different from CV or NLP in any fundamental way. The same recipe that worked for CV and NLP should be able to discover the manifold for robotics and yield a shockingly competent generalist robot. Even if this doesn’t exactly happen, Tedrake points out that attempting to train a large model for general robotics tasks could teach us important things about the manifold of robotics tasks, and perhaps we can leverage this understanding to solve robotics.
• Large models are the best approach we have to get at “common sense” capabilities, which pervade all of robotics: Another thing Russ Tedrake pointed out is that “common sense” pervades almost every robotics task of interest. Consider the task of having a mobile manipulation robot place a mug onto a table. Even if we ignore the challenging problems of finding and localizing the mug, there are a surprising number of subtleties to this problem. What if the table is cluttered and the robot has to move other objects out of the way? What if the mug accidentally falls on the floor and the robot has to pick it up again, re-orient it, and place it on the table? And what if the mug has something in it, so it’s important it’s never overturned? These “edge cases” are actually much more common that it might seem, and often are the difference between success and failure for a task. Moreover, these seem to require some sort of ‘common sense’ reasoning to deal with. Several people argued that large models trained on a large amount of data are the best way we know of to yield some aspects of this ‘common sense’ capability. Thus, they might be the best way we know of to solve general robotics tasks.

As you might imagine, there were a number of arguments against scaling as a practical solution to robotics. Interestingly, almost no one directly disputes that this approach could work in theory. Instead, most arguments fall into one of two buckets: (1) arguing that this approach is simply impractical, and (2) arguing that even if it does kind of work, it won’t really “solve” robotics.

• We currently just don’t have much robotics data, and there’s no clear way we’ll get it: This is the elephant in pretty much every large-scale robot learning room. The Internet is chock-full of data for CV and NLP, but not at all for robotics. Recent efforts to collect very large datasets have required tremendous amounts of time, money, and cooperation, yet have yielded a very small fraction of the amount of vision and text data on the Internet. CV and NLP got so much data because they had an incredible “data flywheel”: tens of millions of people connecting to and using the Internet. Unfortunately for robotics, there seems to be no reason why people would upload a bunch of sensory input and corresponding action pairs. Collecting a very large robotics dataset seems quite hard, and given that we know that a lot of important “emergent” properties only showed up in vision and language models at scale, the inability to get a large dataset could render this scaling approach hopeless.
• Robots have different embodiments: Another challenge with collecting a very large robotics dataset is that robots come in a large variety of different shapes, sizes, and form factors. The output control actions that are sent to a Boston Dynamics Spot robot are very different to those sent to a KUKA iiwa arm. Even if we ignore the problem of finding some kind of common output space for a large trained model, the variety in robot embodiments means we’ll probably have to collect data from each robot type, and that makes the above data-collection problem even harder.
• There is extremely large variance in the environments we want robots to operate in: For a robot to really be “general purpose,” it must be able to operate in any practical environment a human might want to put it in. This means operating in any possible home, factory, or office building it might find itself in. Collecting a dataset that has even just one example of every possible building seems impractical. Of course, the hope is that we would only need to collect data in a small fraction of these, and the rest will be handled by generalization. However, we don’t know how much data will be required for this generalization capability to kick in, and it very well could also be impractically large.
• Training a model on such a large robotics dataset might be too expensive/energy-intensive: It’s no secret that training large foundation models is expensive, both in terms of money and in energy consumption. GPT-4V — OpenAI’s biggest foundation model at the time of this writing — reportedly cost over US $100 million and 50 million KWh of electricity to train. This is well beyond the budget and resources that any academic lab can currently spare, so a larger robotics foundation model would need to be trained by a company or a government of some kind. Additionally, depending on how large both the dataset and model itself for such an endeavor are, the costs may balloon by another order-of-magnitude or more, which might make it completely infeasible.

• The 99.X problem and long tails: Vincent Vanhoucke of Google Robotics started a talk with a provocative assertion: Most — if not all — robot learning approaches cannot be deployed for any practical task. The reason? Real-world industrial and home applications typically require 99.X percent or higher accuracy and reliability. What exactly that means varies by application, but it’s safe to say that robot learning algorithms aren’t there yet. Most results presented in academic papers top out at 80 percent success rate. While that might seem quite close to the 99.X percent threshold, people trying to actually deploy these algorithms have found that it isn’t so: getting higher success rates requires asymptotically more effort as we get closer to 100 percent. That means going from 85 to 90 percent might require just as much — if not more — effort than going from 40 to 80 percent. Vincent asserted in his talk that getting up to 99.X percent is a fundamentally different beast than getting even up to 80 percent, one that might require a whole host of new techniques beyond just scaling.
  • Existing big models don’t get to 99.X percent even in CV and NLP: As impressive and capable as current large models like GPT-4V and DETIC are, even they don’t achieve 99.X percent or higher success rate on previously-unseen tasks. Current robotics models are very far from this level of performance, and I think it’s safe to say that the entire robot learning community would be thrilled to have a general model that does as well on robotics tasks as GPT-4V does on NLP tasks. However, even if we had something like this, it wouldn’t be at 99.X percent, and it’s not clear that it’s possible to get there by scaling either.
• Self-driving car companies have tried this approach, and it doesn’t fully work (yet): This is closely related to the above point, but important and subtle enough that I think it deserves to stand on its own. A number of self-driving car companies — most notably Tesla and Wayve — have tried training such an end-to-end big model on large amounts of data to achieve Level 5 autonomy. Not only do these companies have the engineering resources and money to train such models, but they also have the data. Tesla in particular has a fleet of over 100,000 cars deployed in the real world that it is constantly collecting and then annotating data from. These cars are being teleoperated by experts, making the data ideal for large-scale supervised learning. And despite all this, Tesla has so far been unable to produce a Level 5 autonomous driving system. That’s not to say their approach doesn’t work at all. It competently handles a large number of situations — especially highway driving — and serves as a useful Level 2 (i.e., driver assist) system. However, it’s far from 99.X percent performance. Moreover, data seems to suggest that Tesla’s approach is faring far worse than Waymo or Cruise, which both use much more modular systems. While it isn’t inconceivable that Tesla’s approach could end up catching up and surpassing its competitors performance in a year or so, the fact that it hasn’t worked yet should serve as evidence perhaps that the 99.X percent problem is hard to overcome for a large-scale ML approach. Moreover, given that self-driving is a special case of general robotics, Tesla’s case should give us reason to doubt the large-scale model approach as a full solution to robotics, especially in the medium term.
• Many robotics tasks of interest are quite long-horizon: Accomplishing any task requires taking a number of correct actions in sequence. Consider the relatively simple problem of making a cup of tea given an electric kettle, water, a box of tea bags, and a mug. Success requires pouring the water into the kettle, turning it on, then pouring the hot water into the mug, and placing a tea-bag inside it. If we want to solve this with a model trained to output motor torque commands given pixels as input, we’ll need to send torque commands to all 7 motors at around 40 Hz. Let’s suppose that this tea-making task requires 5 minutes. That requires 7 * 40 * 60 * 5 = 84,000 correct torque commands. This is all just for a stationary robot arm; things get much more complicated if the robot is mobile, or has more than one arm. It is well-known that error tends to compound with longer-horizons for most tasks. This is one reason why — despite their ability to produce long sequences of text — even LLMs cannot yet produce completely coherent novels or long stories: small deviations from a true prediction over time tend to add up and yield extremely large deviations over long-horizons. Given that most, if not all robotics tasks of interest require sending at least thousands, if not hundreds of thousands, of torques in just the right order, even a fairly well-performing model might really struggle to fully solve these robotics tasks.

Okay, now that we’ve sketched out all the main points on both sides of the debate, I want to spend some time diving into a few related points. Many of these are responses to the above points on the ‘against’ side, and some of them are proposals for directions to explore to help overcome the issues raised.

One point that gets brought up a lot against learning-based approaches is the lack of theoretical guarantees. At the time of this writing, we know very little about neural network theory: we don’t really know why they learn well, and more importantly, we don’t have any guarantees on what values they will output in different situations. On the other hand, most classical control and planning approaches that are widely used in robotics have various theoretical guarantees built-in. These are generally quite useful when certifying that systems are safe.

However, there seemed to be general consensus amongst a number of CoRL speakers that this point is perhaps given more significance than it should. Sergey Levine pointed out that most of the guarantees from controls aren’t really that useful for a number of real-world tasks we’re interested in. As he put it: “self-driving car companies aren’t worried about controlling the car to drive in a straight line, but rather about a situation in which someone paints a sky onto the back of a truck and drives in front of the car,” thereby confusing the perception system. Moreover, Scott Kuindersma of Boston Dynamics talked about how they’re deploying RL-based controllers on their robots in production, and are able to get the confidence and guarantees they need via rigorous simulation and real-world testing. Overall, I got the sense that while people feel that guarantees are important, and encouraged researchers to keep trying to study them, they don’t think that the lack of guarantees for learning-based systems means that they cannot be deployed robustly.

In one of the organized debates, Emo Todorov pointed out that existing successful ML systems, like Codex and ChatGPT, work well only because a human interacts with and sanitizes their output. Consider the case of coding with Codex: it isn’t intended to directly produce runnable, bug-free code, but rather to act as an intelligent autocomplete for programmers, thereby making the overall human-machine team more productive than either alone. In this way, these models don’t have to achieve the 99.X percent performance threshold, because a human can help correct any issues during deployment. As Emo put it: “humans are forgiving, physics is not.”

Chelsea Finn responded to this by largely agreeing with Emo. She strongly agreed that all successfully-deployed and useful ML systems have humans in the loop, and so this is likely the setting that deployed robot learning systems will need to operate in as well. Of course, having a human operate in the loop with a robot isn’t as straightforward as in other domains, since having a human and robot inhabit the same space introduces potential safety hazards. However, it’s a useful setting to think about, especially if it can help address issues brought on by the 99.X percent problem.

A number of people at the conference were thinking about creative ways to overcome the real-world data bottleneck without actually collecting more real world data. Quite a few of these people argued that fast, realistic simulators could be vital here, and there were a number of works that explored creative ways to train robot policies in simulation and then transfer them to the real world. Another set of people argued that we can leverage existing vision, language, and video data and then just ‘sprinkle in’ some robotics data. Google’s recent RT-2 model showed how taking a large model trained on internet scale vision and language data, and then just fine-tuning it on a much smaller set robotics data can produce impressive performance on robotics tasks. Perhaps through a combination of simulation and pretraining on general vision and language data, we won’t actually have to collect too much real-world robotics data to get scaling to work well for robotics tasks.

As with any debate, there were quite a few people advocating the middle path. Scott Kuindersma of Boston Dynamics titled one of his talks “Let’s all just be friends: model-based control helps learning (and vice versa)”. Throughout his talk, and the subsequent debates, his strong belief that in the short to medium term, the best path towards reliable real-world systems involves combining learning with classical approaches. In her keynote speech for the conference, Andrea Thomaz talked about how such a hybrid system — using learning for perception and a few skills, and classical SLAM and path-planning for the rest — is what powers a real-world robot that’s deployed in tens of hospital systems in Texas (and growing!). Several papers explored how classical controls and planning, together with learning-based approaches can enable much more capability than any system on its own. Overall, most people seemed to argue that this ‘middle path’ is extremely promising, especially in the short to medium term, but perhaps in the long-term either pure learning or an entirely different set of approaches might be best.

If you’ve read this far, chances are that you’re interested in some set of takeaways/conclusions. Perhaps you’re thinking “this is all very interesting, but what does all this mean for what we as a community should do? What research problems should I try to tackle?” Fortunately for you, there seemed to be a number of interesting suggestions that had some consensus on this.

Despite the various arguments against scaling solving robotics outright, most people seem to agree that scaling in robot learning is a promising direction to be investigated. Even if it doesn’t fully solve robotics, it could lead to a significant amount of progress on a number of hard problems we’ve been stuck on for a while. Additionally, as Russ Tedrake pointed out, pursuing this direction carefully could yield useful insights about the general robotics problem, as well as current learning algorithms and why they work so well.

Even the most vocal proponents of the scaling approach were clear that they don’t think everyone should be working on this. It’s likely a bad idea for the entire robot learning community to put its eggs in the same basket, especially given all the reasons to believe scaling won’t fully solve robotics. Classical robotics techniques have gotten us quite far, and led to many successful and reliable deployments: pushing forward on them or integrating them with learning techniques might be the right way forward, especially in the short to medium terms.

Vincent Vanhoucke made an observation that most papers at CoRL this year were limited to tabletop manipulation settings. While there are plenty of hard tabletop problems, things generally get a lot more complicated when the robot — and consequently its camera view — moves. Vincent speculated that it’s easy for the community to fall into a local minimum where we make a lot of progress that’s specific to the tabletop setting and therefore not generalizable. A similar thing could happen if we work predominantly in simulation. Avoiding these local minima by working on real-world mobile manipulation seems like a good idea.

Separately, Sergey Levine observed that a big reason why LLM’s have seen so much excitement and adoption is because they’re extremely easy to use: especially by non-experts. One doesn’t have to know about the details of training an LLM, or perform any tough setup, to prompt and use these models for their own tasks. Most robot learning approaches are currently far from this. They often require significant knowledge of their inner workings to use, and involve very significant amounts of setup. Perhaps thinking more about how to make robot learning systems easier to use and widely applicable could help improve adoption and potentially scalability of these approaches.

There seemed to be a broadly-held complaint that many robot learning approaches don’t adequately report negative results, and this leads to a lot of unnecessary repeated effort. Additionally, perhaps patterns might emerge from consistent failures of things that we expect to work but don’t actually work well, and this could yield novel insight into learning algorithms. There is currently no good incentive for researchers to report such negative results in papers, but most people seemed to be in favor of designing one.

There were a few people who pointed out that all current approaches — be they learning-based or classical — are unsatisfying in a number of ways. There seem to be a number of drawbacks with each of them, and it’s very conceivable that there is a completely different set of approaches that ultimately solves robotics. Given this, it seems useful to try think outside the box. After all, every one of the current approaches that’s part of the debate was only made possible because the few researchers that introduced them dared to think against the popular grain of their times.

Acknowledgements: Huge thanks to Tom Silver and Leslie Kaelbling for providing helpful comments, suggestions, and encouragement on a previous draft of this post.

—
1 In fact, this was the topic of a popular debate hosted at a workshop on the first day; many of the points in this post were inspired by the conversation during that debate.

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

Enjoy today’s videos!

[ Naver ]

If you’re gonna take a robot dog with you on a hike, the least it could do is carry your backpack for you.

[ Deep Robotics ]

Obligatory reminder that phrases like “no teleoperation” without any additional context can mean many different things.

[ Astribot ]

[ TUM ]

Thanks Berthold!

Hands of blue? Not a good look.

[ Synaptic ]

With all the humanoid stuff going on, there really should be more emphasis on intentional contact—humans lean and balance on things all the time, and robots should too!

[ Inria ]

[ LimX Dynamics ]

[ Science ]

[ TED ]

Greetings from the IEEE International Conference on Robotics and Automation (ICRA) in Yokohama, Japan! We hope you’ve been enjoying our short videos on TikTok, YouTube, and Instagram. They are just a preview of our in-depth ICRA coverage, and over the next several weeks we’ll have lots of articles and videos for you. In today’s edition of Video Friday, we bring you a dozen of the most interesting projects presented at the conference.

Enjoy today’s videos, and stay tuned for more ICRA posts!

Upcoming robotics events for the next few months:

Please send us your events for inclusion.

The following two videos are part of the “ Cooking Robotics: Perception and Motion Planning” workshop, which explored “the new frontiers of ‘robots in cooking,’ addressing various scientific research questions, including hardware considerations, key challenges in multimodal perception, motion planning and control, experimental methodologies, and benchmarking approaches.” The workshop featured robots handling food items like cookies, burgers, and cereal, and the two robots seen in the videos below used knives to slice cucumbers and cakes. You can watch all workshop videos here.

“SliceIt!: Simulation-Based Reinforcement Learning for Compliant Robotic Food Slicing,” by Cristian C. Beltran-Hernandez, Nicolas Erbetti, and Masashi Hamaya from OMRON SINIC X Corporation, Tokyo, Japan.

“Cafe Robot: Integrated AI Skillset Based on Large Language Models,” by Jad Tarifi, Nima Asgharbeygi, Shuhei Takamatsu, and Masataka Goto from Integral AI in Tokyo, Japan, and Mountain View, Calif., USA.

“Autonomous Overhead Powerline Recharging for Uninterrupted Drone Operations,” by Viet Duong Hoang, Frederik Falk Nyboe, Nicolaj Haarhøj Malle, and Emad Ebeid from University of Southern Denmark, Odense, Denmark.

“Learning Quadrupedal Locomotion With Impaired Joints Using Random Joint Masking,” by Mincheol Kim, Ukcheol Shin, and Jung-Yup Kim from Seoul National University of Science and Technology, Seoul, South Korea, and Robotics Institute, Carnegie Mellon University, Pittsburgh, Pa., USA.

“Synthesizing Robust Walking Gaits via Discrete-Time Barrier Functions With Application to Multi-Contact Exoskeleton Locomotion,” by Maegan Tucker, Kejun Li, and Aaron D. Ames from Georgia Institute of Technology, Atlanta, Ga., and California Institute of Technology, Pasadena, Calif., USA.

“Supernumerary Robotic Limbs to Support Post-Fall Recoveries for Astronauts,” by Erik Ballesteros, Sang-Yoep Lee, Kalind C. Carpenter, and H. Harry Asada from MIT, Cambridge, Mass., USA, and Jet Propulsion Laboratory, California Institute of Technology, Pasadena, Calif., USA.

“ArrayBot: Reinforcement Learning for Generalizable Distributed Manipulation through Touch,” by Zhengrong Xue, Han Zhang, Jingwen Cheng, Zhengmao He, Yuanchen Ju, Changyi Lin, Gu Zhang, and Huazhe Xu from Tsinghua Embodied AI Lab, IIIS, Tsinghua University; Shanghai Qi Zhi Institute; Shanghai AI Lab; and Shanghai Jiao Tong University, Shanghai, China.

“SKT-Hang: Hanging Everyday Objects via Object-Agnostic Semantic Keypoint Trajectory Generation,” by Chia-Liang Kuo, Yu-Wei Chao, and Yi-Ting Chen from National Yang Ming Chiao Tung University, in Taipei and Hsinchu, Taiwan, and NVIDIA.

“TEXterity: Tactile Extrinsic deXterity,” by Antonia Bronars, Sangwoon Kim, Parag Patre, and Alberto Rodriguez from MIT and Magna International Inc.

“Out of Sight, Still in Mind: Reasoning and Planning about Unobserved Objects With Video Tracking Enabled Memory Models,” by Yixuan Huang, Jialin Yuan, Chanho Kim, Pupul Pradhan, Bryan Chen, Li Fuxin, and Tucker Hermans from University of Utah, Salt Lake City, Utah, Oregon State University, Corvallis, Ore., and NVIDIA, Seattle, Wash., USA.

“Open Sourse Underwater Robot: Easys,” by Michikuni Eguchi, Koki Kato, Tatsuya Oshima, and Shunya Hara from University of Tsukuba and Osaka University, Japan.

“Sensorized Soft Skin for Dexterous Robotic Hands,” by Jana Egli, Benedek Forrai, Thomas Buchner, Jiangtao Su, Xiaodong Chen, and Robert K. Katzschmann from ETH Zurich, Switzerland, and Nanyang Technological University, Singapore.

It’s hard to think of a more dramatic way to make an entrance than falling from the sky. While it certainly happens often enough on the silver screen, whether or not it can be done in real life is a tantalizing challenge for our entertainment robotics team at Disney Research.

Falling is tricky for two reasons. The first and most obvious is what Douglas Adams referred to as “the sudden stop at the end.” Every second of free fall means another 9.8 m/s of velocity, and that can quickly add up to an extremely difficult energy dissipation problem. The other tricky thing about falling, especially for terrestrial animals like us, is that our normal methods for controlling our orientation disappear. We are used to relying on contact forces between our body and the environment to control which way we’re pointing. In the air, there’s nothing to push on except the air itself!

Finding a solution to these problems is a big, open-ended challenge. In the clip below, you can see one approach we’ve taken to start chipping away at it.

The video shows a small, stick-like robot with an array of four ducted fans attached to its top. The robot has a piston-like foot that absorbs the impact of a small fall, and then the ducted fans keep the robot standing by counteracting any tilting motion using aerodynamic thrust.

Raphael Pilon [left] and Marcela de los Rios evaluate the performance of the monopod balancing robot.Disney Research

The standing portion demonstrates that pushing on the air isn’t only useful during freefall. Conventional walking and hopping robots depend on ground contact forces to maintain the required orientation. These forces can ramp up quickly because of the stiffness of the system, necessitating high bandwidth control strategies. Aerodynamic forces are relatively soft, but even so, they were sufficient to keep our robots standing. And since these forces can also be applied during the flight phase of running or hopping, this approach might lead to robots that run before they walk. The thing that defines a running gait is the existence of a “flight phase” - a time when none of the feet are in contact with the ground. A running robot with aerodynamic control authority could potentially use a gait with a long flight phase. This would shift the burden of the control effort to mid-flight, simplifying the leg design and possibly making rapid bipedal motion more tractable than a moderate pace.

Richard Landon uses a test rig to evaluate the thrust profile of a ducted fan.Disney Research

In the next video, a slightly larger robot tackles a much more dramatic fall, from 65’ in the air. This simple machine has two piston-like feet and a similar array of ducted fans on top. The fans not only stabilize the robot upon landing, they also help keep it oriented properly as it falls. Inside each foot is a plug of single-use compressible foam. Crushing the foam on impact provides a nice, constant force profile, which maximizes the amount of energy dissipated per inch of contraction.

In the case of this little robot, the mechanical energy dissipation in the pistons is less than the total energy needed to be dissipated from the fall, so the rest of the mechanism takes a pretty hard hit. The size of the robot is an advantage in this case, because scaling laws mean that the strength-to-weight ratio is in its favor.

The strength of a component is a function of its cross-sectional area, while the weight of a component is a function of its volume. Area is proportional to length squared, while volume is proportional to length cubed. This means that as an object gets smaller, its weight becomes relatively small. This is why a toddler can be half the height of an adult but only a fraction of that adult’s weight, and why ants and spiders can run around on long, spindly legs. Our tiny robots take advantage of this, but we can’t stop there if we want to represent some of our bigger characters.

Louis Lambie and Michael Lynch assemble an early ducted fan test platform. The platform was mounted on guidewires and was used for lifting capacity tests.Disney Research

In most aerial robotics applications, control is provided by a system that is capable of supporting the entire weight of the robot. In our case, being able to hover isn’t a necessity. The clip below shows an investigation into how much thrust is needed to control the orientation of a fairly large, heavy robot. The robot is supported on a gimbal, allowing it to spin freely. At the extremities are mounted arrays of ducted fans. The fans don’t have enough force to keep the frame in the air, but they do have a lot of control authority over the orientation.

Complicated robots are less likely to survive unscathed when subjected to the extremely high accelerations of a direct ground impact, as you can see in this early test that didn’t quite go according to plan.

In this last video, we use a combination of the previous techniques and add one more capability – a dramatic mid-air stop. Ducted fans are part of this solution, but the high-speed deceleration is principally accomplished by a large water rocket. Then the mechanical legs only have to handle the last ten feet of dropping acceleration.

Whether it’s using water or rocket fuel, the principle underlying a rocket is the same – mass is ejected from the rocket at high speed, producing a reaction force in the opposite direction via Newton’s third law. The higher the flow rate and the denser the fluid, the more force is produced. To get a high flow rate and a quick response time, we needed a wide nozzle that went from closed to open cleanly in a matter of milliseconds. We designed a system using a piece of copper foil and a custom punch mechanism that accomplished just that.

Grant Imahara pressurizes a test tank to evaluate an early valve prototype [left]. The water rocket in action - note the laminar, two-inch-wide flow as it passes through the specially designed nozzleDisney Research

Once the water rocket has brought the robot to a mid-air stop, the ducted fans are able to hold it in a stable hover about ten feet above the deck. When they cut out, the robot falls again and the legs absorb the impact. In the video, the robot has a couple of loose tethers attached as a testing precaution, but they don’t provide any support, power, or guidance.

“It might not be so obvious as to what this can be directly used for today, but these rough proof-of-concept experiments show that we might be able to work within real-world physics to do the high falls our characters do on the big screen, and someday actually stick the landing,” explains Tony Dohi, the project lead.

There are still a large number of problems for future projects to address. Most characters have legs that bend on hinges rather than compress like pistons, and don’t wear a belt made of ducted fans. Beyond issues of packaging and form, making sure that the robot lands exactly where it intends to land has interesting implications for perception and control. Regardless, we think we can confirm that this kind of entrance has–if you’ll excuse the pun–quite the impact.

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

Enjoy today’s videos!

Festo has robot bees!

It’s a very clever design, but the size makes me terrified of whatever the bees are that Festo seems to be familiar with.

[ Festo ]

Boing, boing, boing!

[ USC ]

Why the heck would you take the trouble to program a robot to make sweet potato chips and then not scarf them down yourself?

[ Dino Robotics ]

[ DART Lab ]

Turns out that working on a collaborative task with a robot can make humans less efficient, because we tend to overestimate the robot’s capabilities.

[ CHI 2024 ]

Wing posts a video with the title “What Do Wing’s Drones Sound Like” but only includes a brief snippet—though nothing without background room noise—revealing to curious viewers and listeners exactly what Wing’s drones sound like.

Because, look, a couple seconds of muted audio underneath a voiceover is in fact not really answering the question.

[ Wing ]

[ Michigan Robotics ]

This Microsoft Future Leaders in Robotics and AI Seminar is from Catie Cuan at Stanford, on “Choreorobotics: Teaching Robots How to Dance With Humans.”

[ UMD ]

For years, Shadow Robot Company’s Shadow Hand has arguably been the gold standard for robotic manipulation. Beautiful and expensive, it is able to mimic the form factor and functionality of human hands, which has made it ideal for complex tasks. I’ve personally experienced how amazing it is to use Shadow Hands in a teleoperation context, and it’s hard to imagine anything better.

The problem with the original Shadow hand was (and still is) fragility. In a research environment, this has been fine, except that research is changing: Roboticists no longer carefully program manipulation tasks by, uh, hand. Now it’s all about machine learning, in which you need robotic hands to massively fail over and over again until they build up enough data to understand how to succeed.

“We’ve aimed for robustness and performance over anthropomorphism and human size and shape.” —Rich Walker, Shadow Robot Company

Doing this with a Shadow Hand was just not realistic, which Google DeepMind understood five years ago when it asked Shadow Robot to build it a new hand with hardware that could handle the kind of training environments that now typify manipulation research. So Shadow Robot spent the last half-decade-ish working on a new, three-fingered Shadow Hand, which the company unveiled today. The company is calling it, appropriately enough, “the new Shadow Hand.”

As you can see, this thing is an absolute beast. Shadow Robot says that the new hand is “robust against a significant amount of misuse, including aggressive force demands, abrasion and impacts.” Part of the point, though, is that what robot-hand designers might call “misuse,” robot-manipulation researchers might very well call “progress,” and the hand is designed to stand up to manipulation research that pushes the envelope of what robotic hardware and software are physically capable of.

Shadow Robot understands that despite its best engineering efforts, this new hand will still occasionally break (because it’s a robot and that’s what robots do), so the company designed it to be modular and easy to repair. Each finger is its own self-contained unit that can be easily swapped out, with five Maxon motors in the base of the finger driving the four finger joints through cables in a design that eliminates backlash. The cables themselves will need replacement from time to time, but it’s much easier to do this on the new Shadow Hand than it was on the original. Shadow Robot says that you can swap out an entire New Hand’s worth of cables in the same time it would take you to replace a single cable on the old hand.

Shadow Robot

The new Shadow Hand itself is somewhat larger than a typical human hand, and heavier too: Each modular finger unit weighs 1.2 kilograms, and the entire three-fingered hand is just over 4 kg. The fingers have humanlike kinematics, and each joint can move up to 180 degrees per second with the capability of exerting at least 8 newtons of force at each fingertip. Both force control and position control are available, and the entire hand runs Robot Operating System, the Open Source Robotics Foundation’s collection of open-source software libraries and tools.

One of the coolest new features of this hand is the tactile sensing. Shadow Robot has decided to take the optical route with fingertip sensors, GelSight-style. Each fingertip is covered in soft, squishy gel with thousands of embedded particles. Cameras in the fingers behind the gel track each of those particles, and when the fingertip touches something, the particles move. Based on that movement, the fingertips can very accurately detect the magnitude and direction of even very small forces. And there are even more sensors on the insides of the fingers too, with embedded Hall effect sensors to help provide feedback during grasping and manipulation tasks.

Shadow Robot

The most striking difference here is how completely different of a robotic-manipulation philosophy this new hand represents for Shadow Robot. “We’ve aimed for robustness and performance over anthropomorphism and human size and shape,” says Rich Walker, director of Shadow Robot Company. “There’s a very definite design choice there to get something that really behaves much more like an optimized manipulator rather than a humanlike hand.”

Walker explains that Shadow Robot sees two different approaches to manipulation within the robotics community right now: There’s imitation learning, where a human does a task and then a robot tries to do the task the same way, and then there’s reinforcement learning, where a robot tries to figure out how do the task by itself. “Obviously, this hand was built from the ground up to make reinforcement learning easy.”

The hand was also built from the ground up to be rugged and repairable, which had a significant effect on the form factor. To make the fingers modular, they have to be chunky, and trying to cram five of them onto one hand was just not practical. But because of this modularity, Shadow Robot could make you a five-fingered hand if you really wanted one. Or a two-fingered hand. Or (and this is the company’s suggestion, not mine) “a giant spider.” Really, though, it’s probably not useful to get stuck on the form factor. Instead, focus more on what the hand can do. In fact, Shadow Robot tells me that the best way to think about the hand in the context of agility is as having three thumbs, not three fingers, but Walker says that “if we describe it as that, people get confused.”

There’s still definitely a place for the original anthropomorphic Shadow Hand, and Shadow Robot has no plans to discontinue it. “It’s clear that for some people anthropomorphism is a deal breaker, they have to have it,” Walker says. “But for a lot of people, the idea that they could have something which is really robust and dexterous and can gather lots of data, that’s exciting enough to be worth saying okay, what can we do with this? We’re very interested to find out what happens.”

The Shadow New Hand is available now, starting at about US $74,000 depending on configuration.

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

Enjoy today’s videos!

[ CMU ]

Thanks, Changyi!

[ Silicon Synapse Lab ]

Okay, but where that costume has eyes is not where Spot has eyes, so the Spot in the costume can’t see, right? And now I’m skeptical of the authenticity of the mutual snoot-boop.

[ Boston Dynamics ]

Here’s some video of Field AI’s robots operating in relatively complex and unstructured environments without prior maps. Make sure to read our article from this week for details!

[ Field AI ]

Is it just me, or is it kind of wild that researchers are now publishing papers comparing their humanoid controller to the “manufacturer’s” humanoid controller? It’s like humanoids are a commodity now or something.

[ OSU ]

I, too, am packing armor for ICRA.

[ Pollen Robotics ]

[ MAB Robotics ]

Thanks, Jakub!

[ HEBI ]

Thanks, Dave!

This is how all spills happen at my house too: super passive-aggressively.

[ 1X ]

[ EPFL ]

[ Agility ]

[ ABB ]

I have it on good authority that getting robots to mow a lawn (like, any lawn) is much harder than it looks, but Electric Sheep has built a business around it.

[ Electric Sheep ]

[ Stanford HAI ]

This week’s CMU RI Seminar is from Dieter Fox at NVIDIA and UW, on “Where’s RobotGPT?”

[ CMU ]

One of the biggest challenges for robotics right now is practical autonomous operation in unstructured environments. That is, doing useful stuff in places your robot hasn’t been before and where things may not be as familiar as your robot might like. Robots thrive on predictability, which has put some irksome restrictions on where and how they can be successfully deployed.

But over the past few years, this has started to change, thanks in large part to a couple of pivotal robotics challenges put on by DARPA. The DARPA Subterranean Challenge ran from 2018 to 2021, putting mobile robots through a series of unstructured underground environments. And the currently ongoing DARPA RACER program tasks autonomous vehicles with navigating long distances off-road. Some extremely impressive technology has been developed through these programs, but there’s always a gap between this cutting-edge research and any real-world applications.

Now, a bunch of the folks involved in these challenges, including experienced roboticists from NASA, DARPA, Google DeepMind, Amazon, and Cruise (to name just a few places) are applying everything that they’ve learned to enable real-world practical autonomy for mobile robots at a startup called Field AI.

Field AI was cofounded by Ali Agha, who previously was the leader of NASA JPL’s Aerial Mobility Group. While at JPL, Agha led Team CoSTAR, which won the DARPA Subterranean Challenge Urban Circuit. Agha has also been the principal investigator for DARPA RACER, first with JPL, and now continuing with Field AI. “Field AI is not just a startup,” Agha tells us. “It’s a culmination of decades of experience in AI and its deployment in the field.”

Unstructured environments are where things are constantly changing, which can play havoc with robots that rely on static maps.

The “field” part in Field AI is what makes Agha’s startup unique. Robots running Field AI’s software are able to handle unstructured, unmapped environments without reliance on prior models, GPS, or human intervention. Obviously, this kind of capability was (and is) of interest to NASA and JPL, which send robots to places where there are no maps, GPS doesn’t exist, and direct human intervention is impossible.

But DARPA SubT demonstrated that similar environments can be found on Earth, too. For instance, mines, natural caves, and the urban underground are all extremely challenging for robots (and even for humans) to navigate. And those are just the most extreme examples: robots that need to operate inside buildings or out in the wilderness have similar challenges understanding where they are, where they’re going, and how to navigate the environment around them.

An autonomous vehicle drives across kilometers of desert with no prior map, no GPS, and no road.Field AI

Despite the difficulty that robots have operating in the field, this is an enormous opportunity that Field AI hopes to address. Robots have already proven their worth in inspection contexts, typically where you either need to make sure that nothing is going wrong across a large industrial site, or for tracking construction progress inside a partially completed building. There’s a lot of value here because the consequences of something getting messed up are expensive or dangerous or both, but the tasks are repetitive and sometimes risky and generally don’t require all that much human insight or creativity.

Where Field AI differs from other robotics companies offering these services, as Agha explains, is that his company wants to do these tasks without first having a map that tells the robot where to go. In other words, there’s no lengthy setup process, and no human supervision, and the robot can adapt to changing and new environments. Really, this is what full autonomy is all about: going anywhere, anytime, without human interaction. “Our customers don’t need to train anything,” Agha says, laying out the company’s vision. “They don’t need to have precise maps. They press a single button, and the robot just discovers every corner of the environment.” This capability is where the DARPA SubT heritage comes in. During the competition, DARPA basically said, “here’s the door into the course. We’re not going to tell you anything about what’s back there or even how big it is. Just go explore the whole thing and bring us back the info we’ve asked for.” Agha’s Team CoSTAR did exactly that during the competition, and Field AI is commercializing this capability.

“With our robots, our aim is for you to just deploy it, with no training time needed. And then we can just leave the robots.” —Ali Agha, Field AI

The other tricky thing about these unstructured environments, especially construction environments, is that things are constantly changing, which can play havoc with robots that rely on static maps. “We’re one of the few, if not the only company that can leave robots for days on continuously changing construction sites with minimal supervision,” Agha tells us. “These sites are very complex—every day there are new items, new challenges, and unexpected events. Construction materials on the ground, scaffolds, forklifts, and heavy machinery moving all over the place, nothing you can predict.”

Field AI

Field AI’s approach to this problem is to emphasize environmental understanding over mapping. Agha says that essentially, Field AI is working towards creating “field foundation models” (FFMs) of the physical world, using sensor data as an input. You can think of FFMs as being similar to the foundation models of language, music, and art that other AI companies have created over the past several years, where ingesting a large amount of data from the Internet enables some level of functionality in a domain without requiring specific training for each new situation. Consequently, Field AI’s robots can understand how to move in the world, rather than just where to move. “We look at AI quite differently from what’s mainstream,” Agha explains. “We do very heavy probabilistic modeling.” Much more technical detail would get into Field AI’s IP, says Agha, but the point is that real-time world modeling becomes a by-product of Field AI’s robots operating in the world rather than a prerequisite for that operation. This makes the robots fast, efficient, and resilient.

Developing field-foundation models that robots can use to reliably go almost anywhere requires a lot of real-world data, which Field AI has been collecting at industrial and construction sites around the world for the past year. To be clear, they’re collecting the data as part of their commercial operations—these are paying customers that Field AI has already. “In these job sites, it can traditionally take weeks to go around a site and map where every single target of interest that you need to inspect is,” explains Agha. “But with our robots, our aim is for you to just deploy it, with no training time needed. And then we can just leave the robots. This level of autonomy really unlocks a lot of use cases that our customers weren’t even considering, because they thought it was years away.” And the use cases aren’t just about construction or inspection or other areas where we’re already seeing autonomous robotic systems, Agha says. “These technologies hold immense potential.”

There’s obviously demand for this level of autonomy, but Agha says that the other piece of the puzzle that will enable Field AI to leverage a trillion dollar market is the fact that they can do what they do with virtually any platform. Fundamentally, Field AI is a software company—they make sensor payloads that integrate with their autonomy software, but even those payloads are adjustable, ranging from something appropriate for an autonomous vehicle to something that a drone can handle.

Heck, if you decide that you need an autonomous humanoid for some weird reason, Field AI can do that too. While the versatility here is important, according to Agha, what’s even more important is that it means you can focus on platforms that are more affordable, and still expect the same level of autonomous performance, within the constraints of each robot’s design, of course. With control over the full software stack, integrating mobility with high-level planning, decision making, and mission execution, Agha says that the potential to take advantage of relatively inexpensive robots is what’s going to make the biggest difference toward Field AI’s commercial success.

Same brain, lots of different robots: the Field AI team’s foundation models can be used on robots big, small, expensive, and somewhat less expensive.Field AI

Field AI is already expanding its capabilities, building on some of its recent experience with DARPA RACER by working on deploying robots to inspect pipelines for tens of kilometers and to transport materials across solar farms. With revenue coming in and a substantial chunk of funding, Field AI has even attracted interest from Bill Gates. Field AI’s participation in RACER is ongoing, under a sort of subsidiary company for federal projects called Offroad Autonomy, and in the meantime its commercial side is targeting expansion to “hundreds” of sites on every platform it can think of, including humanoids.

The blistering late-afternoon wind ripped across Camp Taji, a sprawling U.S. military base just north of Baghdad. In a desolate corner of the outpost, where the feared Iraqi Republican Guard had once manufactured mustard gas, nerve agents, and other chemical weapons, a group of American soldiers and Marines were solemnly gathered around an open grave, dripping sweat in the 114-degree heat. They were paying their final respects to Boomer, a fallen comrade who had been an indispensable part of their team for years. Just days earlier, he had been blown apart by a roadside bomb.

As a bugle mournfully sounded the last few notes of “Taps,” a soldier raised his rifle and fired a long series of volleys—a 21-gun salute. The troops, which included members of an elite army unit specializing in explosive ordnance disposal (EOD), had decorated Boomer posthumously with a Bronze Star and a Purple Heart. With the help of human operators, the diminutive remote-controlled robot had protected American military personnel from harm by finding and disarming hidden explosives.

Boomer was a Multi-function Agile Remote-Controlled robot, or MARCbot, manufactured by a Silicon Valley company called Exponent. Weighing in at just over 30 pounds, MARCbots look like a cross between a Hollywood camera dolly and an oversized Tonka truck. Despite their toylike appearance, the devices often leave a lasting impression on those who work with them. In an online discussion about EOD support robots, one soldier wrote, “Those little bastards can develop a personality, and they save so many lives.” An infantryman responded by admitting, “We liked those EOD robots. I can’t blame you for giving your guy a proper burial, he helped keep a lot of people safe and did a job that most people wouldn’t want to do.”

A Navy unit used a remote-controlled vehicle with a mounted video camera in 2009 to investigate suspicious areas in southern Afghanistan.Mass Communication Specialist 2nd Class Patrick W. Mullen III/U.S. Navy

But while some EOD teams established warm emotional bonds with their robots, others loathed the machines, especially when they malfunctioned. Take, for example, this case described by a Marine who served in Iraq:

Some battle-hardened warriors treat remote-controlled devices like brave, loyal, intelligent pets, while others describe them as clumsy, stubborn clods. Either way, observers have interpreted these accounts as unsettling glimpses of a future in which men and women ascribe personalities to artificially intelligent war machines.

Some battle-hardened warriors treat remote-controlled devices like brave, loyal, intelligent pets, while others describe them as clumsy, stubborn clods.

From this perspective, what makes robot funerals unnerving is the idea of an emotional slippery slope. If soldiers are bonding with clunky pieces of remote-controlled hardware, what are the prospects of humans forming emotional attachments with machines once they’re more autonomous in nature, nuanced in behavior, and anthropoid in form? And a more troubling question arises: On the battlefield, will Homo sapiens be capable of dehumanizing members of its own species (as it has for centuries), even as it simultaneously humanizes the robots sent to kill them?

As I’ll explain, the Pentagon has a vision of a warfighting force in which humans and robots work together in tight collaborative units. But to achieve that vision, it has called in reinforcements: “trust engineers” who are diligently helping the Department of Defense (DOD) find ways of rewiring human attitudes toward machines. You could say that they want more soldiers to play “Taps” for their robot helpers and fewer to delight in blowing them up.

For the better part of a decade, several influential Pentagon officials have relentlessly promoted robotic technologies, promising a future in which “humans will form integrated teams with nearly fully autonomous unmanned systems, capable of carrying out operations in contested environments.”

Soldiers test a vertical take-off-and-landing drone at Fort Campbell, Ky., in 2020.U.S. Army

As The New York Times reported in 2016: “Almost unnoticed outside defense circles, the Pentagon has put artificial intelligence at the center of its strategy to maintain the United States’ position as the world’s dominant military power.” The U.S. government is spending staggering sums to advance these technologies: For fiscal year 2019, the U.S. Congress was projected to provide the DOD with US $9.6 billion to fund uncrewed and robotic systems—significantly more than the annual budget of the entire National Science Foundation.

Arguments supporting the expansion of autonomous systems are consistent and predictable: The machines will keep our troops safe because they can perform dull, dirty, dangerous tasks; they will result in fewer civilian casualties, since robots will be able to identify enemies with greater precision than humans can; they will be cost-effective and efficient, allowing more to get done with less; and the devices will allow us to stay ahead of China, which, according to some experts, will soon surpass America’s technological capabilities.

Former U.S. deputy defense secretary Robert O. Work has argued for more automation within the military. Center for a New American Security

Among the most outspoken advocate of a roboticized military is Robert O. Work, who was nominated by President Barack Obama in 2014 to serve as deputy defense secretary. Speaking at a 2015 defense forum, Work—a barrel-chested retired Marine Corps colonel with the slight hint of a drawl—described a future in which “human-machine collaboration” would win wars using big-data analytics. He used the example of Lockheed Martin’s newest stealth fighter to illustrate his point: “The F-35 is not a fighter plane, it is a flying sensor computer that sucks in an enormous amount of data, correlates it, analyzes it, and displays it to the pilot on his helmet.”

The beginning of Work’s speech was measured and technical, but by the end it was full of swagger. To drive home his point, he described a ground combat scenario. “I’m telling you right now,” Work told the rapt audience, “10 years from now if the first person through a breach isn’t a friggin’ robot, shame on us.”

“The debate within the military is no longer about whether to build autonomous weapons but how much independence to give them,” said a 2016 New York Times article. The rhetoric surrounding robotic and autonomous weapon systems is remarkably similar to that of Silicon Valley, where charismatic CEOs, technology gurus, and sycophantic pundits have relentlessly hyped artificial intelligence.

For example, in 2016, the Defense Science Board—a group of appointed civilian scientists tasked with giving advice to the DOD on technical matters—released a report titled “Summer Study on Autonomy.” Significantly, the report wasn’t written to weigh the pros and cons of autonomous battlefield technologies; instead, the group assumed that such systems will inevitably be deployed. Among other things, the report included “focused recommendations to improve the future adoption and use of autonomous systems [and] example projects intended to demonstrate the range of benefits of autonomy for the warfighter.”

The author’s book, War Virtually, is a critical look at how the U.S. military is weaponizing technology and data.University of California Press

Early in the 20th century, military and intelligence agencies began developing robotic systems, which were mostly devices remotely operated by human controllers. But microchips, portable computers, the Internet, smartphones, and other developments have supercharged the pace of innovation. So, too, has the ready availability of colossal amounts of data from electronic sources and sensors of all kinds. The Financial Times reports: “The advance of artificial intelligence brings with it the prospect of robot-soldiers battling alongside humans—and one day eclipsing them altogether.” These transformations aren’t inevitable, but they may become a self-fulfilling prophecy.

All of this raises the question: What exactly is a “robot-soldier”? Is it a remote-controlled, armor-clad box on wheels, entirely reliant on explicit, continuous human commands for direction? Is it a device that can be activated and left to operate semiautonomously, with a limited degree of human oversight or intervention? Is it a droid capable of selecting targets (using facial-recognition software or other forms of artificial intelligence) and initiating attacks without human involvement? There are hundreds, if not thousands, of possible technological configurations lying between remote control and full autonomy—and these differences affect ideas about who bears responsibility for a robot’s actions.

The U.S. military’s experimental and actual robotic and autonomous systems include a vast array of artifacts that rely on either remote control or artificial intelligence: aerial drones; ground vehicles of all kinds; sleek warships and submarines; automated missiles; and robots of various shapes and sizes—bipedal androids, quadrupedal gadgets that trot like dogs or mules, insectile swarming machines, and streamlined aquatic devices resembling fish, mollusks, or crustaceans, to name a few.

Members of a U.S. Air Force squadron test out an agile and rugged quadruped robot from Ghost Robotics in 2023.Airman First Class Isaiah Pedrazzini/U.S. Air Force

The transitions projected by military planners suggest that servicemen and servicewomen are in the midst of a three-phase evolutionary process, which begins with remote-controlled robots, in which humans are “in the loop,” then proceeds to semiautonomous and supervised autonomous systems, in which humans are “on the loop,” and then concludes with the adoption of fully autonomous systems, in which humans are “out of the loop.” At the moment, much of the debate in military circles has to do with the degree to which automated systems should allow—or require—human intervention.

“Ten years from now if the first person through a breach isn’t a friggin’ robot, shame on us.” —Robert O. Work

In recent years, much of the hype has centered around that second stage: semiautonomous and supervised autonomous systems that DOD officials refer to as “human-machine teaming.” This idea suddenly appeared in Pentagon publications and official statements after the summer of 2015. The timing probably wasn’t accidental; it came at a time when global news outlets were focusing attention on a public backlash against lethal autonomous weapon systems. The Campaign to Stop Killer Robots was launched in April 2013 as a coalition of nonprofit and civil society organizations, including the International Committee for Robot Arms Control, Amnesty International, and Human Rights Watch. In July 2015, the campaign released an open letter warning of a robotic arms race and calling for a ban on the technologies. Cosigners included world-renowned physicist Stephen Hawking, Tesla founder Elon Musk, Apple cofounder Steve Wozniak, and thousands more.

In November 2015, Work gave a high-profile speech on the importance of human-machine teaming, perhaps hoping to defuse the growing criticism of “killer robots.” According to one account, Work’s vision was one in which “computers will fly the missiles, aim the lasers, jam the signals, read the sensors, and pull all the data together over a network, putting it into an intuitive interface humans can read, understand, and use to command the mission”—but humans would still be in the mix, “using the machine to make the human make better decisions.” From this point forward, the military branches accelerated their drive toward human-machine teaming.

But there was a problem. Military experts loved the idea, touting it as a win-win: Paul Scharre, in his book Army of None: Autonomous Weapons and the Future of War, claimed that “we don’t need to give up the benefits of human judgment to get the advantages of automation, we can have our cake and eat it too.” However, personnel on the ground expressed—and continue to express—deep misgivings about the side effects of the Pentagon’s newest war machines.

The difficulty, it seems, is humans’ lack of trust. The engineering challenges of creating robotic weapon systems are relatively straightforward, but the social and psychological challenges of convincing humans to place their faith in the machines are bewilderingly complex. In high-stakes, high-pressure situations like military combat, human confidence in autonomous systems can quickly vanish. The Pentagon’s Defense Systems Information Analysis Center Journal noted that although the prospects for combined human-machine teams are promising, humans will need assurances:

In a 2015 article, U.S. Navy Commander Greg Smith provided a candid assessment of aviators’ distrust in aerial drones. After describing how drones are often intentionally separated from crewed aircraft, Smith noted that operators sometimes lose communication with their drones and may inadvertently bring them perilously close to crewed airplanes, which “raises the hair on the back of an aviator’s neck.” He concluded:

U.S. Marines [top] prepare to launch and operate a MQ-9A Reaper drone in 2021. The Reaper [bottom] is designed for both high-altitude surveillance and destroying targets.Top: Lance Cpl. Gabrielle Sanders/U.S. Marine Corps; Bottom: 1st Lt. John Coppola/U.S. Marine Corps

Yet Pentagon leaders place an almost superstitious trust in those systems, and seem firmly convinced that a lack of human confidence in autonomous systems can be overcome with engineered solutions. In a commentary, Courtney Soboleski, a data scientist employed by the military contractor Booz Allen Hamilton, makes the case for mobilizing social science as a tool for overcoming soldiers’ lack of trust in robotic systems.

Soon, the wary warfighter will likely be subjected to new forms of training that focus on building trust between robots and humans. Already, robots are being programmed to communicate in more human ways with their users for the explicit purpose of increasing trust. And projects are currently underway to help military robots report their deficiencies to humans in given situations, and to alter their functionality according to the machine’s perceived emotional state of the user.

At the DEVCOM Army Research Laboratory, military psychologists have spent more than a decade on human experiments related to trust in machines. Among the most prolific is Jessie Chen, who joined the lab in 2003. Chen lives and breathes robotics—specifically “agent teaming” research, a field that examines how robots can be integrated into groups with humans. Her experiments test how humans’ lack of trust in robotic and autonomous systems can be overcome—or at least minimized.

For example, in one set of tests, Chen and her colleagues deployed a small ground robot called an Autonomous Squad Member that interacted and communicated with infantrymen. The researchers varied “situation-awareness-based agent transparency”—that is, the robot’s self-reported information about its plans, motivations, and predicted outcomes—and found that human trust in the robot increased when the autonomous “agent” was more transparent or honest about its intentions.

The Army isn’t the only branch of the armed services researching human trust in robots. The U.S. Air Force Research Laboratory recently had an entire group dedicated to the subject: the Human Trust and Interaction Branch, part of the lab’s 711th Human Performance Wing, located at Wright-Patterson Air Force Base, in Ohio.

In 2015, the Air Force began soliciting proposals for “research on how to harness the socio-emotional elements of interpersonal team/trust dynamics and inject them into human-robot teams.” Mark Draper, a principal engineering research psychologist at the Air Force lab, is optimistic about the prospects of human-machine teaming: “As autonomy becomes more trusted, as it becomes more capable, then the Airmen can start off-loading more decision-making capability on the autonomy, and autonomy can exercise increasingly important levels of decision-making.”

Air Force researchers are attempting to dissect the determinants of human trust. In one project, they examined the relationship between a person’s personality profile (measured using the so-called Big Five personality traits: openness, conscientiousness, extraversion, agreeableness, neuroticism) and his or her tendency to trust. In another experiment, entitled “Trusting Robocop: Gender-Based Effects on Trust of an Autonomous Robot,” Air Force scientists compared male and female research subjects’ levels of trust by showing them a video depicting a guard robot. The robot was armed with a Taser, interacted with people, and eventually used the Taser on one. Researchers designed the scenario to create uncertainty about whether the robot or the humans were to blame. By surveying research subjects, the scientists found that women reported higher levels of trust in “Robocop” than men.

The issue of trust in autonomous systems has even led the Air Force’s chief scientist to suggest ideas for increasing human confidence in the machines, ranging from better android manners to robots that look more like people, under the principle that

It’s impossible to know the degree to which the trust engineers will succeed in achieving their objectives. For decades, military trainers have trained and prepared newly enlisted men and women to kill other people. If specialists have developed simple psychological techniques to overcome the soldier’s deeply ingrained aversion to destroying human life, is it possible that someday, the warfighter might also be persuaded to unquestioningly place his or her trust in robots?

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

Enjoy today’s videos!

[ DARPA ]

[ Github ]

[ Sanctuary AI ]

[ Zipline ]

Work those hips!

[ RoMeLa ]

This thing is kind of terrifying, and I’m fascinated by it.

[ AVFL ]

From Humanoids 2018.

[ Paper ] via [ JSK Lab ]

Thanks, Kento!

Ballbots are uniquely capable of pushing wheelchairs—arguably better than legged platforms, because they can move in any direction without having to reposition themselves.

[ Paper ]

[ Charge Robotics ]

Robots doing precision assembly is cool and all, but those vibratory bowl sorters seem like magic.

[ FANUC ]

[ QUT ]

[ CMU ]

Here’s a little more video of that robot elevator, but I’m wondering why it’s so slow—clamp those bots in there and rocket that elevator up and down!

[ NAVER ]

[ Northwestern ]

There’s something so soothing about watching a lithium battery get wrecked and burn for 8 minutes.

[ Hardcore Robotics ]

[ JPL ]

[ TED ]

What’s a secret to getting more students to participate in an IEEE society? Give them a seat at the table so they have a say in how the organization is run.

That’s what the IEEE Robotics and Automation Society has done. Budding engineers serve on the RAS board of directors, have voting privileges, and work within technical committees.

“They have been given a voice in how the society runs because, in the end, students are among the main beneficiaries,” says Enrica Tricomi, chair of the RAS’s student activities committee. The SAC is responsible for student programs and benefits. It also makes recommendations to the society’s board about new offerings.

The IEEE Robotics and Automation Society isn’t focused only on boosting its student membership. It also wants to get more young people interested in pursuing a robotics career. One way the society’s volunteers try to inspire the next generation of roboticists is through IEEE Spectrum’s award-winning Robots website. The interactive guide features more than 250 real-world robots, with thousands of photos, videos, and exclusive interactives, plus news and detailed technical specifications.

The site is designed for anyone interested in robotics, including expert and beginner enthusiasts, researchers, entrepreneurs, students, STEM educators, and other teachers.

Schools and students across the globe use the site. Volunteers on the RAS steering committee suggest robots to add, and they help support new content creation on the site.

“You feel listened to and valued whenever there are official decisions to be made, because the board also wants to know the perspective of students on how to offer benefits to the RAS members, especially for young researchers, since hopefully they will be the society’s future leaders,” says Tricomi, a bioengineer who is pursuing a Ph.D. in robotics at Heidelberg University, in Germany.

The society’s approach has paid off. Since 2018, student membership has grown by more than 50 percent to 5,436. The number of society chapters at student branches has increased from 312 in 2021 to 450.

The ability to express opinions isn’t the only reason students are joining, Tricomi says. The society recently launched several programs to engage them, including career fairs, travel grants, and networking opportunities with researchers.

As SAC chair, Tricomi is a voting member of RAS’s administrative committee, which oversees the society’s operations. She says having voting privileges shows “how important it is to the society to have student representation.”

“We receive a lot of support from the highest levels of the society, specifically the society president, Aude Billard, and past president Frank Chongwoo Park,” Tricomi says. “RAS boards have been rejuvenated to engage students even more and represent their voices. The chairs of these boards—including technical activities, conference activities, and publication activities—want to know the SAC chair and cochairs’ opinion on whether the new activities are benefiting students.”

Student members now can serve on IEEE technical committees that involve robotics in the role of student representatives.

That was an initiative from Kyujin Cho, IEEE Technical Activities vice president. Tricomi says the designation benefits young engineers because they learn about ongoing research in their field and because they have direct access to researchers.

Student representatives also help organize conference workshops.

The students had a hand in creating a welcome kit for conference attendees. The initiative, led by Amy Kyungwon Han, Technical Activities associate vice president, lists each day’s activities and their location.

“I think that all of us, especially those who are younger, can actively contribute and make a difference not only for the society and for ourselves but also for our peers.”

Being engaged with the technical topic in which the students work provides them with career growth, visibility in their field, and an opportunity to share their point of view with peers, Tricomi says.

“Being young, the first time that you express your opinion in public, you always feel uncomfortable because you don’t have much experience,” she says. “This is the opposite of the message the society wants to send. We want to listen to students’ voices because they are an important part of the society.”

Tricomi herself recently became a member of the Technical Activities board.

She joined, she says, because “this is kind of a technical family by choice. And you want to be active and contribute to your family, right? I think that all of us, especially those who are younger, can actively contribute and make a difference not only for the society and for ourselves but also for our peers.”

Several new initiatives have been rolled out at the society’s flagship conferences. The meetings have always included onsite events for students to network with each other and to mingle with researchers over lunch. The events give the budding engineers an opportunity to talk with leaders they normally wouldn’t meet, Tricomi says.

“It’s much appreciated, especially by very young or shy students,” she says.

Some luncheons have included sessions on career advice from leaders in academia and industry, or from startup founders—giving the students a sense of what it’s like to work for such organizations.

Conferences now include career fairs, where students can meet with hiring companies.

The society also developed a software platform that allows candidates to upload their résumé onsite. If they are a match for an open position, interviews can be held on the spot.

A variety of travel grants have been made available to students with limited resources so they can present their research papers at the society’s major conferences. More than 200 travel grants were awarded to the 2023 IEEE International Conference on Robotics and Automation, Tricomi says.

“It’s very important for them to be there, presenting their work, gaining visibility, sharing their research, and also networking,” she says.

The new IDEA (inclusion, diversity, equity, and accessibility) travel grant for underrepresented groups was established by the society’s IEEE Women in Engineering committee and its chair, Karinne Ramirez Amaro. The grant can help students who are not presenters to attend conferences. It also helps increase diversity within the robotics field, Tricomi says.

The Member Support Program is a new initiative from the RAS member activities board’s vice president, Katja Mombaur, and past vice president Stefano Stramigioli. Financial support to attend the annual International Conference on Intelligent Robots and Systems is available to members and students who have contributed to the society’s mission-related activities. The projects include organizing workshops, discussions, lectures, or networking events at conferences or sponsored events; serving on boards or committees; or writing papers that were accepted for publication by conferences or journals.

The society also gets budding engineers involved in publication activities through its Young Reviewers Program, which introduces them to best practices for peer review. Senior reviewers assign the students papers to check and oversee their work.

Tricomi joined the society in 2021 shortly after starting her Ph.D. program at Heidelberg. Her research is in wearable assistive robotics for human augmentation or rehabilitation purposes. She holds a master’s degree in biomedical engineering from Politecnico di Torino, in Italy.

She was new to the field of robotics, so her Ph.D. advisor, IEEE Senior Member Lorenzo Masia, encouraged her to volunteer for the society. She is now transitioning to the role of SAC senior chair, and she says she is eager to collaborate with the new team to promote student and early career engagement within the robotics field.

“I’ve realized I’ve grown up a lot in the two years since I started as chair,” she says. “At the beginning, I was much shier. I really want my colleagues to experience the same personal and professional growth as I have. You learn not only technical skills but also soft skills, which are very important in your career.”

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

Enjoy today’s videos!

[ SpaceHopper ]

It’s still kind of wild to me that it’s now possible to just build a robot like Menteebot. Having said that, at present it looks to be a fairly long way from being able to usefully do tasks in a reliable way.

[ Menteebot ]

Look, it’s the robot we all actually want!

[ Github ]

I wasn’t quite sure what made this building especially “robot-friendly” until I saw the DEDICATED ROBOT ELEVATOR.

[ NAVER ]

[ LimX Dynamics ]

[ Paper ]

Thanks to its “buttock-contact sensors,” JSK’s musculoskeletal humanoid has mastered(ish) the chair-scoot.

[ University of Tokyo ]

Thanks, Kento!

Physical therapy seems like a great application for a humaonid robot when you don’t really need that humanoid robot to do much of anything.

[ Fourier Intelligence ]

[ JPL ]

No thank you!

[ Paper ]

[ Mitsubishi ]

Thanks, Devesh!

[ Agility ]

The PR2 just refuses to die. Last time I checked, official support for it ceased in 2016!

[ University of Bremen ]

[ DARPA ]

Sometimes robots that exist for one single purpose that they only do moderately successfully while trying really hard are the best of robots.

[ CMU ]

The paper delves into the significance of Secure Runtime Assurance (SRTA) for the operational integrity and safety of autonomous robotics groups, with a focus on drones. It presents a comprehensive view of how SRTA has evolved from traditional runtime assurance methods to address the dynamic and complex nature of autonomous systems. Through integrating artificial intelligence and machine learning, SRTA seeks to tackle the multifaceted challenges autonomous systems face, highlighting the need for adaptive, scalable, and secure solutions. Emphasizing a hierarchical approach to decision-making, the paper also highlights the critical role of redundancy in ensuring reliability and anticipates future advancements in RTA technologies. This paper reflects an ongoing effort to harmonize safety and efficiency within regulatory frameworks for autonomous robotics.

Stephen Cass: Hello and welcome to Fixing the Future, an IEEE Spectrum podcast where we look at concrete solutions to tough problems. I’m your host, Stephen Cass, a senior editor at IEEE Spectrum. And before I start, I just want to tell you that you can get the latest coverage of some of Spectrum’s most important beats, including AI, climate change, and robotics, by signing up for one of our free newsletters. Just go to spectrum.ieee.org/newsletters to subscribe. We’ve been covering the drone delivery company Zipline in Spectrum for several years, and I do encourage listeners to check out our great onsite reporting from Rwanda in 2019 when we visited one of Zipline’s dispatch centers for delivering vital medical supplies into rural areas. But now it’s 2024, and Zipline is expanding into commercial drone delivery in the United States, including into urban areas, and hitting some recent milestones. Here to talk about some of those milestones today, we have Keenan Wyrobek, Zipline’s co-founder and CTO. Keenan, welcome to the show.

Keenan Wyrobek: Great to be here. Thanks for having me.

Cass: So before we get into what’s going on with the United States, can you first catch us up on how things have been going on with Rwanda and the other African countries you’ve been operating in?

Wyrobek: Yeah, absolutely. So we’re now operating in eight countries, including here in the US. That includes a handful of countries in Africa, as well as Japan and Europe. So in Africa, it’s really exciting. So the scale is really impressive, basically. As we’ve been operating, started eight years ago with blood, then moved into vaccine delivery and delivering many other things in the healthcare space, as well as outside the healthcare space. We can talk a little bit about in things like animal husbandry and other things. The scale is really what’s exciting. We have a single distribution center there that now regularly flies more than the equivalent of once the equator of the Earth every day. And that’s just from one of a whole bunch of distribution centers. That’s where we are really with that operation today.

Cass: So could you talk a little bit about those non-medical systems? Because this was very much how we’d seen blood being parachuted down from these drones and reaching those distant centers. What other things are you delivering there?

Wyrobek: Yeah, absolutely. So start with blood, like you said, then vaccines. We’ve now done delivered well over 15 million vaccine doses, lots of other pharmaceutical use cases to hospitals and clinics, and more recently, patient home delivery for chronic care of things like hypertension, HIV-positive patients, and things like that. And then, yeah, moved into some really exciting use cases and things like animal husbandry. One that I’m personally really excited about is supporting these genetic diversity campaigns. It’s one of those things very unglamorous, but really impactful. One of the main sources of protein around the world is cow’s milk. And it turns out the difference between a non-genetically diverse cow and a genetically diverse cow can be 10x difference in milk production. And so one of the things we deliver is bull semen. We’re very good at the cold chain involved in that as we’ve mastered in vaccines and blood. And that’s just one of many things we’re doing in other spaces outside of healthcare directly.

Cass: Oh, fascinating. So turning now to the US, it seems like there’s been two big developments recently. One is you’re getting close to deploying Platform 2, which has some really fascinating tech that allows packages to be delivered very precisely by tether. And I do want to talk about that later. But first, I want to talk about a big milestone you had late last year. And this was something that goes by the very unlovely acronym of a BVLOS flight. Can you tell us what a BVLOS stands for and why that flight was such a big deal?

Wryobek: Yeah, “beyond visual line of sight.” And so that is basically, before this milestone last year, all drone deliveries, all drone operations in the US were done by people standing on the ground, looking at the sky, that line of sight. And that’s how basically we made sure that the drones were staying clear of aircraft. This is true of everybody. Now, this is important because in places like the United States, many aircraft don’t and aren’t required to carry a transponder, right? So transponders where they have a radio signal that they’re transmitting their location that our drones can listen to and use to maintain separation. And so the holy grail of basically scalable drone operations, of course, it’s physically impossible to have people standing around all the world staring at the sky, and is a sensing solution where you can sense those aircraft and avoid those aircraft. And this is something we’ve been working on for a long time and got the approval for late last year with the FAA, the first-ever use of sensors to detect and avoid for maintaining safety in the US airspace, which is just really, really exciting. That’s now been in operations in two distribution centers here, one in Utah and one in Arkansas ever since.

Cass: So could you just tell us a little bit about how that tech works? It just seems to be quite advanced to trust a drone to recognize, “Oh, that is an actual airplane that’s a Cessna that’s going to be here in about two minutes and is a real problem,” or, “No, it’s a hawk, which is just going about his business and I’m not going to ever come close to it at all because it’s so far away.

Wryobek: Yeah, this is really fun to talk about. So just to start with what we’re not doing, because most people expect us to use either a radar for this or cameras for this. And basically, those don’t work. And the radar, you would need such a heavy radar system to see 360 degrees all the way around your drone. And this is really important because two things to kind of plan in your mind. One is we’re not talking about autonomous driving where cars are close together. Aircraft never want to be as close together as cars are on a road, right? We’re talking about maintaining hundreds of meters of separation, and so you sense it a long distance. And drones don’t have right of way. So what that means is even if a plane’s coming up behind the drone, you got to sense that plane and get out of the way. And so to have enough radar on your drone that you can actually see far enough to maintain that separation in every direction, you’re talking about something that weighs many times the weight of a drone and it just doesn’t physically close. And so we started there because that’s sort of where we assumed and many people assume that’s the place to start. Then looked at cameras. Cameras have lots of drawbacks. And fundamentally, you can sort of-- we’ve all had this, you taken your phone and tried to take a picture of an airplane and you look at the picture, you can’t see the airplane. Yeah. It takes so many pixels of perfectly clean lenses to see an aircraft at a kilometer or two away that it really just is not practical or robust enough. And that’s when we went back to the drawing board and it ended up where we ended up, which is using an array of microphones to listen for aircraft, which works very well at very long distances to then maintain separation from those other aircraft.

Cass: So yeah, let’s talk about Platform 2 a little bit more because I should first explain for listeners who maybe aren’t familiar with Zipline that these are not the kind of the little purely sort of helicopter-like drones. These are these fixed wing with sort of loiter capability and hovering capabilities. So they’re not like your Mavic drones and so on. These have a capacity then for long-distance flight, which is what it gives them.

Wyrobek: Yeah. And maybe to jump into Platform 2— maybe starting with Platform 1, what does it look like? So Platform 1 is what we’ve been operating around the world for years now. And this basically looks like a small airplane, right? In the industry referred to as a fixed-wing aircraft. And it’s fixed wing because to solve the problem of going from a metro area to surrounding countryside, really two things matter. Your range and long range and low cost. And a fixed-wing aircraft over something that can hover has something like an 800% advantage in range and cost. And that’s why we did fix wing because it actually works for our customers for their needs for that use case. Platform 2 is all about, how do you deliver to homes and in metro areas where you need an incredible amount of precision to deliver to nearly every home. And so Platform 2—we call our drone zips—our drone, it flies out to the delivery site. Instead of floating a package down to a customer like Platform 1 does, it hovers. Platform 2 hovers and lowers down what we call a droid. And so the droids on tether. The drone stays way up high, about 100 meters up high, and the drone lowers down. And the drone itself-- sorry, the droid itself, it lowers down, it can fly. Right? So you think of it as like the tether does the heavy lifting, but the droid has fans. So if it gets hit by a gust of wind or whatnot, it can still stay very precisely on track and come in and deliver it to a very small area, put the package down, and then be out of there seconds later.

Cass: So let me get this right. Platform 2 is kind of as a combo, fixed wing and rotor wing. It’s like a VTOL like that. I’m cheating here a little bit because my colleague Evan Ackerman has a great Q&A on the Spectrum website with you, some of your team members about the nitty-gritty of how that design was evolved. But first off, it’s like a little droid thing at the end of the tether. How much extra precision do all those fans and stuff give you?

Wyrobek: Oh, massive, right? We can come down and hit a target within a few centimeters of where we want to deliver, which means we can deliver. Like if you have a small back porch, which is really common, right, in a lot of urban areas to have a small back porch or a small place on your roof or something like that, we can still just deliver as long as we have a few feet of open space. And that’s really powerful for being able to serve our customers. And a lot of people think of Platform 2 as like, “Hey, it’s a slightly better way of doing maybe a DoorDash-style operation, people in cars driving around.” And to be clear, it’s not slightly better. It’s massively better, much faster, more environmentally friendly. But we have many contracts for Platform 2 in the health space with US Health System Partners and Health Systems around the world. And what’s powerful about these customers in terms of their needs is they really need to serve all of their customers. And this is where a lot of our sort of-- this is where our engineering effort goes is how do you make a system that doesn’t just kind of work for some folks, and they can use it if they want to, but a health system is like, “No, I want this to work for everybody in my health network.” And so how do we get to that near 100 percent serviceability? And that’s what this droid really enables us to do. And of course, it has all these other magic benefits too. It makes some of the hardest design problems in this space much, much easier. The safety problem gets much easier by keeping the drone way up high.

Cass: Yeah, how high is Platform 2 hovering when it’s doing its deliveries?

Wyrobek: About 100 meters, so 300 plus feet, right? We’re talking about high up as a football field is long. And so it’s way up there. And it also helps with things like noise, right? We don’t want to live in a future where drones are all around us sounding like swarms of insects. We want drones to make no noise. We want them to just melt into the background. And so it makes that kind of problem much easier as well. And then, of course, the droid gets other benefits where for many products, we don’t need any packaging at all. We can just deliver the product right onto a table in your porch. And not just from a cost perspective, but again, from— we’re all familiar with the nightmare of packaging from deliveries we get. Eliminating packaging just has to be our future. And we’re really excited to advance that future.

Cass: From Evan’s Q&A, I know that a lot of effort went into making the droid element look rather adorable. Why was that so important?

Wryobek: Yeah, I like to describe it as sort of a cross between three things, if you kind of picture this, like a miniature little fan boat, right, because it has some fan, a big fan on the back, looks like a little fan boat, combined with sort of a baby seal, combined with a toaster. It sort of has that look to it. And making it adorable, there’s a bunch of sort of human things that matter, right? I want this to be something that when my grandmother, who’s not a tech-savvy, gets these deliveries, it’s approachable. It doesn’t come off as sort of scary. And when you make something cute, not only does it feel approachable, but it also forces you to get the details right so it is approachable, right? The rounded corners, right? This sounds really benign, but a lot of robots, it turns out if you bump into them, they scratch you. And we want you to be able to bump into this droid, and this is no big deal. And so getting the surfaces right, getting them— the surface is made sort of like a helmet foam. If you can picture that, right? The kind of thing you wouldn’t be afraid to touch if it touched you. And so getting it both to be something that feels safe, but is something that actually is safe to be around, those two things just matter a lot. Because again, we’re not designing this for some piloty kind of low-volume thing. Our customers want this in phenomenal volume. And so we really want this to be something that we’re all comfortable around.

Cass: Yeah, and one thing I want to pull out from that Q&A as well is it was an interesting note, because you mentioned it has three fans, but they’re rather unobtrusive. And the original design, you had two big fans on the sides, which was very great for maneuverability. But you had to get rid of those and come up with a three-fan design. And maybe you can explain why that was so.

Wryobek: Yeah, that’s a great detail. So the original design, the picture, it was like, imagine the package in the middle, and then kind of on either side of the package, two fans. So when you looked at it, it kind of looked like— I don’t know. It kind of looked like the package had big mouse ears or something. And when you looked at it, everybody had the same reaction. You kind of took this big step back. It was like, “Whoa, there’s this big thing coming down into my yard.” And when you’re doing this kind of user testing, we always joke, you don’t need to bring users in if it already makes you take a step back. And this is one of those things where like, “That’s just not good enough, right, to even start with that kind of refined design.” But when we got the sort of profile of it smaller, the way we think about it from a design experiment perspective is we want to deliver a large package. So basically, the droid needs to be as sucked down as small additional volume around that package as possible. So we spent a lot of time figuring out, “Okay, how do you do that sort of physically and aesthetically in a way that also gets that amazing performance, right? Because when I say performance, what I’m talking about is we still need it to work when the winds are blowing really hard outside and still can deliver precisely. And so it has to have a lot of aero performance to do that and still deliver precisely in essentially all weather conditions.

Cass: So I guess I just want to ask you then is, what kind of weight and volume are you able to deliver with this level of precision?

Wryobek: Yeah, yeah. So we’ll be working our way up to eight pounds. I say working our way up because that’s part of, once you launch a product like this, there’s refinement you can do overtime on many layers, but eight pounds, which was driven off, again, these health use cases. So it does basically 100 percent of what our health partners need to do. And it turns out it’s, nearly 100 percent of what we want to do in meal delivery. And even in the goods sector, I’m impressed by the percentage of goods we can deliver. One of our partners we work with, we can deliver over 80 percent of what they have in their big box store. And yeah, it’s wildly exceeding expectations on nearly every axis there. And volume, it’s big. It’s bigger than a shoebox. I don’t have a great-- I’m trying to think of a good reference to kind of bring it to life. But it looks like a small cooler basically inside. And it can comfortably fit a meal for four to give you a sense of the amount of food you can fit in there. Yeah.

Cass: So we’ve seen this history of Zipline in rural areas, and now we’re talking about expanding operations in more urban areas, but just how urban? I don’t imagine that we’ll see the zip lines of zooming around, say, the very hemmed-in streets, say, here in Midtown Manhattan. So what level of urban are we talking about?

Wryobek: Yeah, so the way we talk about it internally in our design process is basically we call three-story sprawl. Manhattan is the place where when we think of New York, we’re not talking about Manhattan, but most of the rest of New York, we are talking about it, right? Like the Bronx, things like that. We just have this sort of three stories forever. And that’s a lot of the world out here in California, that’s most of San Francisco. I think it’s something like 98 percent of San Francisco is that. If you’ve ever been to places like India and stuff like that, the cities, it’s just sort of this three stories going for a really long way. And that’s what we’re really focused on. And that’s also where we provide that incredible value because that’s also matches where the hardest traffic situations and things like that can make any other sort of terrestrial on-demand delivery be phenomenally late.

Cass: Well, no, I live out in Queens, so I agree there’s not much skyscrapers out there. Although there are quite a few trees and so on, but at the same time, there’s usually some sort of sidewalk availability. So is that kind of what you’re hoping to get into?

Wyrobek: Exactly. So as long as you’ve got a porch with a view of the sky or an alley with a view of the sky, it can be literally just a few feet, we can get in there, make a delivery, and be on our way.

Cass: And so you’ve done this preliminary test with the FAA, the BVLOS test, and so on. How close do you think you are to, and you’re working with a lot of partners, to really seeing this become routine commercial operations?

Wyrobek: Yeah, yeah. So at relatively limited scale, our operations here in Utah and in Arkansas that are leveraging that FAA approval for beyond visual line-of-sight flight operations, that’s been all day, every day now since our approval last year. With Platform 2, we’re really excited. That’s coming later this year. We’re currently in the phase of basically massive-scale testing. So we now have our production hardware and we’re taking it through a massive ground testing campaign. So this picture dozens of thermal chambers and five chambers and things like that just running to really both validate that we have the reliability we need and flush out any issues that we might have missed so we can address that difference between what we call the theoretical reliability and the actual reliability. And that’s running in parallel to a massive flight test campaign. Same idea, right? We’re slowly ramping up the flight volume as we fly into heavier conditions really to make sure we know the limits of the system. We know its actual reliability and true scaled operations so we can get the confidence that it’s ready to operate for people.

Cass: So you’ve got Platform 2. What’s kind of next on your technology roadmap for any possible platform three?

Wyrobek: Oh, great question. Yeah, I can’t comment on platform three at this time, but. And I will also say, Zipline is pouring our heart into Platform 2 right now. Getting Platform 2 ready for this-- the way I like to talk about this internally is today, we fly about four times the equator of the Earth in our operations on average. And that’s a few thousand flights per day. But the demand we have is for more like millions of flights per day, if not beyond. And so on the log scale, right, we’re halfway there. Three hours of magnitude down, three more zeros to come. And the level of testing, the level of systems engineering, the level of refinement required to do that is a lot. And there’s so many systems from weather forecasting to our onboard autonomy and our fleet management systems. And so to highlight one team, our system test team run by this really impressive individual named Juan Albanell, this team has taken us from where we were two years ago, where we had shown the concept at a very prototype stage of this delivery experience, and we’ve done the first order math kind of on the architecture and things like that through the iterations in test to actually make sure we had a drone that could actually fly in all these weather conditions with all the robustness and tolerance required to actually go to this global scale that Platform 2 is targeting.

Cass: Well, that’s fantastic. Well, I think there’s a lot more to talk about to come up in the future, and we look forward to talking with Zipline again. But for today, I’m afraid we’re going to have to leave it there. But it was really great to have you on the show, Keenan. Thank you so much.

Wyrobek: Cool. Absolutely, Stephen. It was a pleasure to speak with you.

Cass: So today on Fixing the Future, we were talking with Zipline’s Keenan Wyrobek about the progress of commercial drone deliveries. For IEEE Spectrum, I’m Stephen Cass, and I hope you’ll join us next time.

Boston Dynamics has just introduced a new Atlas humanoid robot, replacing the legendary hydraulic Atlas and intended to be a commercial product. This is huge news from the company that has spent the last decade building the most dynamic humanoids that the world has ever seen, and if you haven’t read our article about the announcement (and seen the video!), you should do that right now.

We’ve had about a decade of pent-up questions about an all-electric productized version of Atlas, and we were lucky enough to speak with Boston Dynamics CEO Robert Playter to learn more about where this robot came from and how it’s going to make commercial humanoid robots (finally) happen.

Robert Playter was the Vice President of Engineering at Boston Dynamics starting in 1994, which I’m pretty sure was back when Boston Dynamics still intended to be a modeling and simulation company rather than a robotics company. Playter became the CEO in 2019, helping the company make the difficult transition from R&D to commercial products with Spot, Stretch, and now (or very soon) Atlas.

We talked with Playter about what the heck took Boston Dynamics so long to make this robot, what the vision is for Atlas as a product, all that extreme flexibility, and what comes next.

Robert Playter on:

• What Took So Long
• The Product Approach
• A General Purpose Robot?
• Hydraulic Versus Electric
• Extreme Range of Motion
• Atlas’ Head
• Advantages in Commercialization
• What’s Next

IEEE Spectrum: So what’s going on?

Robert Playter: Boston Dynamics has built an all-electric humanoid. It’s our newest generation of what’s been an almost 15-year effort in developing humanoids. We’re going to launch it as a product, targeting industrial applications, logistics, and places that are much more diverse than where you see Stretch—heavy objects with complex geometry, probably in manufacturing type environments. We’ve built our first robot, and we believe that’s really going to set the bar for the next generation of capabilities for this whole industry.

What took you so long?!

Playter: Well, we wanted to convince ourselves that we knew how to make a humanoid product that can handle a great diversity of tasks—much more so than our previous generations of robots—including at-pace bimanual manipulation of the types of heavy objects with complex geometry that we expect to find in industry. We also really wanted to understand the use cases, so we’ve done a lot of background work on making sure that we see where we can apply these robots fruitfully in industry.

We’ve obviously been working on this machine for a while, as we’ve been doing parallel development with our legacy Atlas. You’ve probably seen some of the videos of Atlas moving struts around—that’s the technical part of proving to ourselves that we can make this work. And then really designing a next generation machine that’s going to be an order of magnitude better than anything the world has seen.

“We’re not anxious to just show some whiz-bang tech, and we didn’t really want to indicate our intent to go here until we were convinced that there is a path to a product.” —Robert Playter, Boston Dynamics

With Spot, it felt like Boston Dynamics developed the product first, without having a specific use case in mind: you put the robot out there and let people discover what it was good for. Is your approach different with Atlas?

Playter: You’re absolutely right. Spot was a technology looking for a product, and it’s taken time for us to really figure out the product market fit that we have in industrial inspection. But the challenge of that experience has left us wiser about really identifying the target applications before you say you’re going to build these things at scale.

Stretch is very different, because it had a clear target market. Atlas is going to be more like Stretch, although it’s going to be way more than a single task robot, which is kind of what Stretch is. Convincing ourselves that we could really generalize with Atlas has taken a little bit of time. This is going to be our third product in about four years. We’ve learned so much, and the world is different from that experience.

[back to top]

Is your vision for Atlas one of a general purpose robot?

Playter: It definitely needs to be a multi-use case robot. I believe that because I don’t think there’s very many examples where a single repetitive task is going to warrant these complex robots. I also think, though, that the practical matter is that you’re going to have to focus on a class of use cases, and really making them useful for the end customer. The lesson we’ve learned with both Spot and Stretch is that it’s critical to get out there and actually understand what makes this robot valuable to customers while making sure you’re building that into your development cycle. And if you can start that before you’ve even launched the product, then you’ll be better off.

[back to top]

How does thinking of this new Atlas as a product rather than a research platform change things?

Playter: I think the research that we’ve done over the past 10 or 15 years has been essential to making a humanoid useful in the first place. We focused on dynamic balancing and mobility and being able to pick something up and still maintain that mobility—those were research topics of the past that we’ve now figured out how to manage and are essential, I think, to doing useful work. There’s still a lot of work to be done on generality, so that humanoids can pick up any one of a thousand different parts and deal with them in a reasonable way. That level of generality hasn’t been proven yet; we think there’s promise, and that AI will be one of the tools that helps solve that. And there’s still a lot of product prototyping and iteration that will come out before we start building massive numbers of these things and shipping them to customers.

“This robot will be stronger at most of its joints than a person, and even an elite athlete, and will have a range of motion that exceeds anything a person can ever do.” —Robert Playter, Boston Dynamics

For a long time, it seemed like hydraulics were the best way of producing powerful dynamic motions for robots like Atlas. Has that now changed?

Playter: We first experimented with that with the launch of Spot. We had the same issue years ago, and discovered that we could build powerful lightweight electric motors that had the same kind of responsiveness and strength, or let’s say sufficient responsiveness and strength, to really make that work. We’ve designed an even newer set of really compact actuators into our electric Atlas, which pack the strength of essentially an elite human athlete into these tiny packages that make an electric humanoid feasible for us. So, this robot will be stronger at most of its joints than a person, and even an elite athlete, and will have a range of motion that exceeds anything a person can ever do. We’ve also compared the strength of our new electric Atlas to our hydraulic Atlas, and the electric Atlas is stronger.

[back to top]

In the context of Atlas’ range of motion, that introductory video was slightly uncomfortable to watch, which I’m sure was deliberate. Why introduce the new Atlas in that way?

Playter: These high range of motion actuators are going to enable a unique set of movements that ultimately will let the robot be very efficient. Imagine being able to turn around without having to take a bunch of steps to turn your whole body instead. The motions we showed [in the video] are ones where our engineers were like, “hey, with these joints, we could get up like this!” And it just wasn’t something we had that really thought about before. This flexibility creates a palette that you can design new stuff on, and we’re already having fun with it and we decided we wanted to share that excitement with the world.

[back to top]

“Everybody will buy one robot—we learned that with Spot. But they won’t start by buying fleets, and you don’t have a business until you can sell multiple robots to the same customer.” —Robert Playter, Boston Dynamics

This does seem like a way of making Atlas more efficient, but I’ve heard from other folks working on humanoids that it’s important for robots to move in familiar and predictable ways for people to be comfortable working around them. What’s your perspective on that?

Playter: I do think that people are going to have to become familiar with our robot; I don’t think that means limiting yourself to human motions. I believe that ultimately, if your robot is stronger or more flexible, it will be able to do things that humans can’t do, or don’t want to do.

One of the real challenges of making a product useful is that you’ve got to have sufficient productivity to satisfy a customer. If you’re slow, that’s hard. We learned that with Stretch. We had two generations of Stretch, and the first generation did not have a joint that let it pivot 180 degrees, so it had to ponderously turn around between picking up a box and dropping it off. That was a killer. And so we decided “nope, gotta have that rotational joint.” It lets Stretch be so much faster and more efficient. At the end of the day, that’s what counts. And people will get used to it.

What can you tell me about the head?

Boston Dynamics CEO Robert Playter said the head on the new Atlas robot has been designed not to mimic the human form but rather “to project something else: a friendly place to look to gain some understanding about the intent of the robot.”Boston Dynamics

Playter: The old Atlas did not have an articulated head. But having an articulated head gives you a tool that you can use to indicate intent, and there are integrated lights which will be able to communicate to users. Some of our original concepts had more of a [human] head shape, but for us they always looked a little bit threatening or dystopian somehow, and we wanted to get away from that. So we made a very purposeful decision about the head shape, and our explicit intent was for it not to be human-like. We’re trying to project something else: a friendly place to look to gain some understanding about the intent of the robot.

The design borrows from some friendly shapes that we’d seen in the past. For example, there’s the old Pixar lamp that everybody fell in love with decades ago, and that informed some of the design for us.

[back to top]

How do you think the decade(s) of experience working on humanoids as well as your experience commercializing Spot will benefit you when it comes to making Atlas into a product?

Playter: This is our third product, and one of the things we’ve learned is that it takes way more than some interesting technology to make a product work. You have to have a real use case, and you have to have real productivity around that use case that a customer cares about. Everybody will buy one robot—we learned that with Spot. But they won’t start by buying fleets, and you don’t have a business until you can sell multiple robots to the same customer. And you don’t get there without all this other stuff—the reliability, the service, the integration.

When we launched Spot as a product several years ago, it was really about transforming the whole company. We had to take on all of these new disciplines: manufacturing, service, measuring the quality and reliability of our robots and then building systems and tools to make them steadily better. That transformation is not easy, but the fact that we’ve successfully navigated through that as an organization means that we can easily bring that mindset and skill set to bear as a company. Honestly, that transition takes two or three years to get through, so all of the brand new startup companies out there who have a prototype of a humanoid working—they haven’t even begun that journey.

There’s also cost. Building something effectively at a reasonable cost so that you can sell it at a reasonable cost and ultimately make some money out of it, that’s not easy either. And frankly, without the support of Hyundai which is of course a world-class manufacturing expert, it would be really challenging to do it on our own.

So yeah, we’re much more sober about what it takes to succeed now. We’re not anxious to just show some whiz-bang tech, and we didn’t really want to indicate our intent to go here until we were convinced that there is a path to a product. And I think ultimately, that will win the day.

[back to top]

What will you be working on in the near future, and what will you be able to share?

Playter: We’ll start showing more of the dexterous manipulation on the new Atlas that we’ve already shown on our legacy Atlas. And we’re targeting proof of technology testing in factories at Hyundai Motor Group [HMG] as early as next year. HMG is really excited about this venture; they want to transform their manufacturing and they see Atlas as a big part of that, and so we’re going to get on that soon.

[back to top]

What do you think other robotics folks will find most exciting about the new Atlas?

Playter: Having a robot with so much power and agility packed into a relatively small and lightweight package. I’ve felt honored in the past that most of these other companies compare themselves to us. They say, “well, where are we on the Boston Dynamics bar?” I think we just raised the bar. And that’s ultimately good for the industry, right? People will go, “oh, wow, that’s possible!” And frankly, they’ll start chasing us as fast as they can—that’s what we’ve seen so far. I think it’ll end up pulling the whole industry forward.

Yesterday, Boston Dynamics bid farewell to the iconic Atlas humanoid robot. Or, the hydraulically-powered version of Atlas, anyway—if you read between the lines of the video description (or even just read the actual lines of the video description), it was pretty clear that although hydraulic Atlas was retiring, it wasn’t the end of the Atlas humanoid program at Boston Dynamics. In fact, Atlas is already back, and better than ever.

Today, Boston Dynamics is introducing a new version of Atlas that’s all-electric. It’s powered by batteries and electric actuators, no more messy hydraulics. It exceeds human performance in terms of both strength and flexibility. And for the first time, Boston Dynamics is calling this humanoid robot a product. We’ll take a look at everything that Boston Dynamics is announcing today, and have even more detail in this Q&A with Boston Dynamics CEO Robert Playter.

Boston Dynamics’ new electric humanoid has been simultaneously one of the worst and best kept secrets in robotics over the last year or so. What I mean is that it seemed obvious, or even inevitable, that Boston Dynamics would take the expertise in humanoids that it developed with Atlas and combine that with its experience productizing a fully electric system like Spot. But just because something seems inevitable doesn’t mean it actually is inevitable, and Boston Dynamics has done an admirable job of carrying on as normal while building a fully electric humanoid from scratch. And here it is:

It’s all new, it’s all electric, and some of those movements make me slightly uncomfortable (we’ll get into that in a bit). The blog post accompanying the video is sparse on technical detail, but let’s go through the most interesting parts:

In 2010, we took a look at all the humanoid robots then in existence. You could, I suppose, argue that Honda was putting real R&D effort into ASIMO back then, but yeah, pretty much all those other humanoid robots came from research rather than industry. Now, it feels like we’re up to our eyeballs in commercial humanoids, but over the past couple of years, as startups have appeared out of nowhere with brand new humanoid robots, Boston Dynamics (to most outward appearances) was just keepin’ on with that R&D. Today’s announcement certainly changes that.

Boston Dynamics

This is a significant advantage for Boston Dynamics—through Hyundai, they can essentially be their own first customer for humanoid robots, offering an immediate use case in a very friendly transitional environment. Tesla has a similar advantage with Optimus, but Boston Dynamics also has experience sourcing and selling and supporting Spot, which are those business-y things that seem like they’re not the hard part until they turn out to actually be the hard part.

World’s most dynamic humanoid, you say? Awesome! Prove it! On video! With outtakes!

Now we’re getting to the good bits. It’s especially notable here that the electric version of Atlas will be “stronger” than the previous hydraulic version, because for a long time hydraulics were really the only way to get the kind of explosively powerful repetitive dynamic motions that enabled Atlas to do jumps and flips. And the switch away from hydraulics enables that extra range of motion now that there aren’t hoses and stuff to deal with.

It’s also pretty clear that the new Atlas is built to continue the kind of work that hydraulic Atlas has been doing, manipulating big and heavy car parts. This is in sharp contrast to most other humanoid robots that we’ve seen, which have primarily focused on moving small objects or bins around in warehouse environments.

This is all par for the course now, but it’s also not particularly meaningful without more information. “We will give our robots new capabilities through machine learning and AI” is what every humanoid robotics company (and most other robotics companies) are saying, but I’m not sure that we’re there yet, because there’s an “okay but how?” that needs to happen first. I’m not saying that it won’t happen, just pointing out that until it does happen, it hasn’t happened.

Agility Robotics has a similar philosophy with Digit, which has a mostly humanoid form factor to operate in human environments but also uses a non-human leg design because Agility believes that it works better. Atlas is a bit more human-like with its overall design, but there are some striking differences, including both range of motion and the head, both of which we’ll be talking more about.

The introductory video with the new Atlas really punches you in the face with this: Atlas is not constrained by human range of motion and will leverage its extra degrees of freedom to operate faster and more efficiently, even if you personally might find some of those motions a little bit unsettling.

Boston Dynamics

As Marco Hutter pointed out, most commercial robots (humanoids included) are really only targeting tasks that are dull, because dull usually means repetitive, and robots are very good at repetitive. Dirty is a little more complicated, and dangerous is a lot more complicated than that. I appreciate that Boston Dynamics is targeting those other categories of tasks from the outset.

There’s a lot more to building a successful robotics company than building a successful robot. Arguably, building a successful robot is not even the hardest part, long term. Having over 1500 Spot robots deployed with customers gives them a well-established product infrastructure baseline to expand from with the new Atlas.

Taking a step back, let’s consider the position that Boston Dynamics is in when it comes to the humanoid space right now.

The new Atlas appears to be a reasonably mature platform with explicit commercial potential, but it’s not yet clear if this particular version of Atlas is truly commercially viable, in terms of being manufacturable and supportable at scale—it’s Atlas 001, after all. There’s likely a huge amount of work that still needs to be done, but it’s a process that the company has already gone through with Spot. My guess is that Boston Dynamics has some catching up to do with respect to other humanoid companies that are already entering pilot projects.

In terms of capabilities, even though the new Atlas hardware is new, it’s not like Boston Dynamics is starting from scratch, since they’re already transferring skills from hydraulic Atlas onto the new platform. But, we haven’t seen the new Atlas doing any practical tasks yet, so it’s hard to tell how far along that is, and it would be premature to assume that hydraulic Atlas doing all kinds of amazing things in YouTube videos implies that electric Atlas can do similar things safely and reliably in a product context. There’s a gap there, possibly an enormous gap, and we’ll need to see more from the new Atlas to understand where it’s at.

And obviously, there’s a lot of competition in humanoids right now, although I’d like to think that the potential for practical humanoid robots to be useful in society is significant enough that there will be room for lots of different approaches. Boston Dynamics was very early to humanoids in general, but they’re somewhat late to this recent (and rather abrupt) humanoid commercialization push. This may not be a problem, especially if Atlas is targeting applications where its strength and flexibility sets it apart from other robots in the space, and if their depth of experience deploying commercial robotic platforms helps them to scale quickly.

Boston Dynamics

An electric Atlas may indeed have been inevitable, and it’s incredibly exciting to (finally!) see Boston Dynamics take this next step towards a commercial humanoid, which would deliver on more than a decade of ambition stretching back through the DARPA Robotics Challenge to PETMAN. We’ve been promised more manipulation footage soon, and Boston Dynamics expects that Atlas will be in the technology demonstration phase in Hyundai factories as early as next year.

We have a lot more questions, but we have a lot more answers, too: you’ll find a Q&A with Boston Dynamics CEO Robert Playter right here.

In a new video posted today, Boston Dynamics is sending off its hydraulic Atlas humanoid robot. “For almost a decade,” the video description reads, “Atlas has sparked our imagination, inspired the next generations of roboticists, and leapt over technical barriers in the field. Now it’s time for our hydraulic Atlas robot to kick back and relax.”

Hydraulic Atlas has certainly earned some relaxation; Boston Dynamics has been absolutely merciless with its humanoid research program. This isn’t a criticism—sometimes being merciless to your hardware is necessary to push the envelope of what’s possible. And as spectators, we just just get to enjoy it, and this highlight reel includes unseen footage of Atlas doing things well along with unseen footage of Atlas doing things not so well. Which, let’s be honest, is what we’re all really here for.

There’s so much more to the history of Atlas than this video shows. Atlas traces its history back to a DARPA project called PETMAN (Protection Ensemble Test Mannequin), which we first wrote about in 2009, so long ago that we had to dig up our own article on the Wayback Machine. As contributor Mikell Taylor wrote back then:

PETMAN is designed to test the suits used by soldiers to protect themselves against chemical warfare agents. It has to be capable of moving just like a soldier—walking, running, bending, reaching, army crawling—to test the suit’s durability in a full range of motion. To really simulate humans as accurately as possible, PETMAN will even be able to “sweat”.

Relative to the other humanoid robots out there at the time (the most famous of which, by far, was Honda’s ASIMO), PETMAN’s movement and balance were very, very impressive. Also impressive was the presumably unintentional way in which this PETMAN video synced up with the music video to Stayin’ Alive by the Bee Gees. Anyway, DARPA was suitably impressed by all this impressiveness, and chose Boston Dynamics to build another humanoid robot to be used for the DARPA Robotics Challenge. That robot was unveiled ten years ago.

The DRC featured a [still looking for a collective noun for humanoid robots] of Atlases, and it seemed like Boston Dynamics was hooked on the form factor, because less than a year after the DRC Finals the company announced the next generation of Atlas, which could do some useful things like move boxes around. Every six months or so, Boston Dynamics put out a new Atlas video, with the robot running or jumping or dancing or doing parkour, leveraging its powerful hydraulics to impress us every single time. There was really nothing like hydraulic Atlas in terms of dynamic performance, and you could argue that there still isn’t. This is a robot that will be missed.

The original rendering of Atlas, followed by four generations of the robot.Boston Dynamics/IEEE Spectrum

Now, if you’re wondering why Boston Dynamics is saying “it’s time for our hydraulic Atlas robot to kick back and relax,” rather than just “our Atlas robot,” and if you’re also wondering why the video description ends with “take a look back at everything we’ve accomplished with the Atlas platform “to date,” well, I can’t help you. Some people might attempt to draw some inferences and conclusions from that very specific and deliberate language, but I would certainly not be one of them, because I’m well known for never speculating about anything.

I would, however, point out a few things that have been obvious for a while now. Namely, that:

• Boston Dynamics has been focusing fairly explicitly on commercialization over the past several years
• Complex hydraulic robots are not product friendly because (among other things) they tend to leave puddles of hydraulic fluid on the carpet
• Boston Dynamics has been very successful with Spot as a productized electric platform based on earlier hydraulic research platforms
• Fully electric commercial humanoids really seems to be where robotics is at right now

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

Enjoy today’s videos!

I think suggesting that robots can’t fall is much less useful than instead suggesting that robots can fall and get quickly and easily get back up again.

[ Deep Robotics ]

Sanctuary AI says that this video shows Phoenix operating at “human-equivalent speed,” but they don’t specify which human or under which conditions. Though it’s faster than I would be, that’s for sure.

[ Sanctuary AI ]

“Suzume” is an animated film by Makoto Shinkai, in which one of the characters gets turned into a three-legged chair:

Shintaro Inoue from JSK Lab at the University of Tokyo has managed to build a robotic version of that same chair, which is pretty impressive:

[ Github ]

Thanks, Shintaro!

[ 1X ]

[ Electric Sheep ]

[ Starship ]

Agility Robotics is doing its best to answer the (very common) question of “Okay, but what can humanoid robots actually do?”

[ Agility Robotics ]

Digit is great and everything, but Cassie will always be one of my favorite robots.

[ CoRIS ]

[ OmniNxt ]

[ Paper ]

[ Paper ]

There is a notable lack of consistency about where exactly Boston Dynamics wants you to think Spot’s eyes are.

[ Boston Dynamics ]

As with every single cooking video, there’s a lot of background prep that’s required for this robot to cook an entire meal, but I would utterly demolish those fries.

[ Dino Robotics ]

Here’s everything you need to know about Wing delivery drones, except for how much human time they actually require and the true cost of making deliveries by drone, because those things aren’t fun to talk about.

[ Wing ]

This CMU Teruko Yata Memorial Lecture is by Agility Robotics’ Jonathan Hurst, on “Human-Centric Robots and How Learning Enables Generality.”

[ CMU ]

We tend to think about hopping robots from the ground up. That is, they start on the ground, and then, by hopping, incorporate a aerial phase into their locomotion. But there’s no reason why aerial robots can’t approach hopping from the other direction, by adding a hopping ground phase to flight. Hopcopter is the first robot that I’ve ever seen give this a try, and it’s remarkably effective, combining a tiny quadrotor with a springy leg to hop hop hop all over the place.

Songnan Bai, Runze Ding, Song Li, and Bingxuan Pu

So why in the air is it worth adding a pogo stick to an otherwise perfectly functional quadrotor? Well, flying is certainly a valuable ability to have, but does take a lot of energy. If you pay close attention to birds (acknowledged experts in the space), they tend to spend a substantial amount of time doing their level best not to fly, often by walking on the ground or jumping around in trees. Not flying most of the time is arguably one of the things that makes birds so successful—it’s that multimodal locomotion capability that has helped them to adapt to so many different environments and situations.

Hopcopter is multimodal as well, although in a slightly more restrictive sense: Its two modes are flying and intermittent flying. But the intermittent flying is very important, because cutting down on that flight phase gives Hopcopter some of the same efficiency benefits that birds experience. By itself, a quadrotor of hopcopter’s size can stay airborne for about 400 seconds, while Hopcopter can hop continuously for more than 20 minutes. If your objective is to cover as much distance as possible, Hopcopter might not be as effective as a legless quadrotor. But if your objective is instead something like inspection or search and rescue, where you need to spend a fair amount of time not moving very much, hopping could be significantly more effective.

Hopcopter is a small quadcopter (specifically a Crazyflie) attached to a springy pogo-stick leg.Songnan Bai, Runze Ding, Song Li, and Bingxuan Pu

Hopcopter can reposition itself on the fly to hop off of different surfaces.Songnan Bai, Runze Ding, Song Li, and Bingxuan Pu

The actual hopping is mostly passive. Hopcopter’s leg is two rigid pieces connected by rubber bands, with a Crazyflie microcopter stapled to the top. During a hop, the Crazyflie can add directional thrust to keep the hops hopping and alter its direction as well as its height, from 0.6 meters to 1.6 meters. There isn’t a lot of room for extra sensors on Hopcopter, but the addition of some stabilizing fins allow for continuous hopping without any positional feedback.

Besides vertical hopping, Hopcopter can also position itself in midair to hop off of surfaces at other orientations, allowing it to almost instantaneously change direction, which is a neat trick.

And it can even do mid air somersaults, because why not?

Hopcopter’s repertoire of tricks includes somersaults.Songnan Bai, Runze Ding, Song Li, and Bingxuan Pu

The researchers, based at the City University of Hong Kong, say that the Hopcopter technology (namely, the elastic leg) could be easily applied to most other quadcopter platforms, turning them into Hopcopters as well. And if you’re more interested in extra payload rather than extra endurance, it’s possible to use hopping in situations where a payload would be too heavy for continuous flight.

The researchers published their work 10 April in Science Robotics.

Last December, the AI Institute announced that it was opening an office in Zurich as a European counterpart to its Boston headquarters and recruited Marco Hutter to helm the office. Hutter also runs the Robotic Systems Lab at ETH Zurich, arguably best known as the origin of the ANYmal quadruped robot (but it also does tons of other cool stuff).

We’re doing our best to keep close tabs on the institute, because it’s one of a vanishingly small number of places that currently exist where roboticists have the kind of long-term resources and vision necessary to make substantial progress on really hard problems that aren’t quite right for either industry or academia. The institute is still scaling up (and the branch in Zurich has only just kicked things off), but we did spot some projects that the Boston folks have been working on, and as you can see from the clips at the top of this page, they’re looking pretty cool.

Meanwhile, we had a chance to check in with Marco Hutter to get a sense of what the Zurich office will be working on and how he’s going to be solving all of the hard problems in robotics. All of them!

How much can you tell us about what you’ll be working on at the AI Institute?

Marco Hutter: If you know the research that I’ve been doing in the past at ETH and with our startups, there’s an overlap on making systems more mobile, making systems more able to interact with the world, making systems in general more capable on the hardware and software side. And that’s what the institute strives for.

The institute describes itself as a research organization that aims to solve the most important and fundamental problems in robotics and AI. What do you think those problems are?

Marco Hutter is the head of the AI Institute’s new Zurich branch.Swiss Robotics Day

Hutter: There are lots of problems. If you’re looking at robots today, we have to admit that they’re still pretty stupid. The way they move, their capability of understanding their environment, the way they’re able to interact with unstructured environments—I think we’re still lacking a lot of skills on the robotic side to make robots useful in all of the tasks we wish them to do. So we have the ambition of having these robots taking over all these dull, dirty, and dangerous jobs. But if we’re honest, today the biggest impact is really only for the dull part. And I think these dirty and dangerous jobs, where we really need support from robots, that’s still going to take a lot of fundamental work on the robotics and AI side to make enough progress for robots to become useful tools.

What is it about the institute that you think will help robotics make more progress in these areas?

Hutter: I think the institute is one of these unique places where we are trying to bring the benefits of the academic world and the benefits from this corporate world together. In academia, we have all kinds of crazy ideas and we try to develop them in all different directions, but at the same time, we have limited engineering support, and we can only go so far. Making robust and reliable hardware systems is a massive effort, and that kind of engineering is much better done in a corporate lab.

You’ve seen this a little bit with the type of work my lab has been doing in the past. We built simple quadrupeds with a little bit of mobility, but in order to make them robust, we eventually had to spin it out. We had to bring it to the corporate world, because for a research group, a pure academic group, it would have been impossible. But at the same time, you’re losing something, right? Once you go into your corporate world and you’re running a business, you have to be very focused; you can’t be that explorative and free anymore.

So if you bring these two things together through the institute, with long-term planning, enough financial support, and brilliant people both in the U.S. and Europe working together, I think that’s what will hopefully help us make significant progress in the next couple of years.

“We’re very different from a traditional company, where at some point you need to have a product that makes money. Here, it’s really about solving problems and taking the next step.” —Marco Hutter, AI Institute

And what will that actually mean in the context of dynamically mobile robots?

Hutter: If you look at Boston Dynamics’ Atlas doing parkour, or ANYmal doing parkour, these are still demonstrations. You don’t see robots running around in the forests or robots working in mines and doing all kinds of crazy maintenance operations, or in industrial facilities, or construction sites, you name it. We need to not only be able to do this once as a prototype demonstration, but to have all the capabilities that bring that together with environmental perception and understanding to make this athletic intelligence more capable and more adaptable to all kinds of different environments. This is not something that from today to tomorrow we’re going to see it being revolutionized—it will be gradual, steady progress because I think there’s still a lot of fundamental work that needs to be done.

I feel like the mobility of legged robots has improved a lot over the last five years or so, and a lot of that progress has come from Boston Dynamics and also from your lab. Do you feel the same?

Hutter: There has always been progress; the question is how much you can zoom in or zoom out. I think one thing has changed quite a bit, and that’s the availability of robotic systems to all kinds of different research groups. If you look back a decade, people had to build their own robots, they had to do the control for the robots, they had to work on the perception for the robots, and putting everything together like that makes it extremely fragile and very challenging to make something that works more than once. That has changed, which allows us to make faster progress.

Marc Raibert (founder of the AI Institute) likes to show videos of mountain goats to illustrate what robots should be (or will be?) capable of. Does that kind of thing inspire you as well?

Hutter: If you look at the animal kingdom, there’s so many things you can draw inspiration from. And a lot of this stuff is not only the cognitive side; it’s really about pairing the cognitive side with the mechanical intelligence of things like the simple-seeming hooves of mountain goats. But they’re really not that simple, they’re pretty complex in how they interact with the environment. Having one of these things and not the other won’t allow the animal to move across its challenging environment. It’s the same thing with the robots.

It’s always been like this in robotics, where you push on the hardware side, and your controls become better, so you hit a hardware limitation. So both things have to evolve hand in hand. Otherwise, you have an over-dimensioned hardware system that you can’t use because you don’t have the right controls, or you have very sophisticated controls and your hardware system can’t keep up.

How do you feel about all of the investment into humanoids right now, when quadrupedal robots with arms have been around for quite a while?

Hutter: There’s a lot of ongoing research on quadrupeds with arms, and the nice thing is that these technologies that are developed for mobile systems with arms are the same technologies that are used in humanoids. It’s not different from a research point of view, it’s just a different form factor for the system. I think from an application point of view, the story from all of these companies making humanoids is that our environment has been adapted to humans quite a bit. A lot of tasks are at the height of a human standing, right? A quadruped doesn’t have the height to see things or to manipulate things on a table. It’s really application dependent, and I wouldn’t say that one system is better than the other.