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.

Why Scaling Might Work

• 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 to 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 “commonsense” 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.

Why Scaling Might Not Work

It’s impractical

• 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.

Even if it works as well as in CV/NLP, it won’t solve robotics

• 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.

Miscellaneous Related Arguments

We can probably deploy learning-based approaches robustly

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.

What if we strive to deploy Human-in-the-Loop systems?

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.

Maybe we don’t need to collect that much real-world data for scaling

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.

Maybe combining classical and learning-based approaches can give us the best of both worlds

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.

What Can/Should We Take Away From All This?

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.

We should pursue the direction of trying to just scale up learning with very large datasets

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.

We should also pursue other existing directions

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.

We should focus more on real-world mobile manipulation and easy-to-use systems

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.

We should be more forthright about things that don’t work

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.

We should try to do something totally new

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.