Chelsea Finn: This is the State of the Art in Robotics

Robots can already fold laundry, make espresso, clean kitchens, and assemble things. The harder problem is getting them to do those tasks reliably, for long periods of time, without a human babysitting them. At Startup School 2026, Physical Intelligence cofounder Chelsea Finn explains what it takes to build general-purpose robots that work in the real world. She shares how reinforcement learning pushed robot throughput up 2x, how their systems can run autonomously for hours, and why she believes robotics is entering its GPT era: moving from specialized models toward general-purpose systems that can work across tasks, robots, and environments. Transcript: https://www.ycrootaccess.com/p/chelsea-finn-on-the-next-decade-in

• 0:06Speaker 00

  Everyone. Today I'm going to be talking about the state of the art of physical intelligence. And in particular, two years ago, I founded a company called Physical Intelligence. And we're really interested in how we can basically develop any robot or allow any robot to do any task in the real world. And I actually spoke at this event a year ago, last year. And at the event last year, I shared some of our progress in at the company at Physical Intelligence, where we could do things, really complicated tasks, like folding, unloading and folding laundry.

• 0:49Speaker 00

  And I also talked about how, for the first time, we showed how robots can do useful tasks in environments and rooms they've never been in before. Now, since then, since one year ago, we have gotten robots to do a lot of other really cool things. So, for example, we've gotten robots to be able to wash a greasy pan in the top right, or peel a carrot in the video below that, or make a grilled cheese sandwich in the video below that, or slice a zucchini, and so forth. But what I'd really like to focus on today isn't cool videos of robots doing lots of different things,

• 1:31Speaker 00

  but what it actually takes to get robots to be useful in the real world. And specifically, how can we develop general-purpose robots that are useful in the real world? Now, there are two aspects of this. The first is general purpose, how we can develop general-purpose models. And the second is actually bringing in those models to the real world so that they can actually have an impact and be useful to people. And in the first part, I'll talk about being useful in the real world. So to actually bring a technology to the real world, I think we need to figure out, it's helpful to actually look at what people have done in the past to bring AI into the real world. And if we look at a timeline of major production launches

• 2:13Speaker 00

  that are leveraging technology like machine learning, we can see a timeline like this. So I think the really, the first early examples of machine learning being used for real in the real world were for things like product recommendations and ad ranking. And then, five years later, we started to see not just machine learning being used, but deep learning being used for the same sorts of applications. And this was a really exciting advance because deep learning is an algorithm that actually isn't, like you can kind of really apply it out of the box to scenarios that involve really complex, like inputs and outputs and so forth, and it makes it easier to translate to other applications.

• 2:54Speaker 00

  But from there, I think that even more exciting kind of moment in time that we saw in terms of machine learning and AI in production was in 2022 with the launch of ChatGPT. And this was the first time where we saw a general-purpose model truly being used by many different people in the real world. Within five days, ChatGPT had reached a million users. And then, of course, more recently, we've seen things like Claude Code also be incredibly useful, hopefully, to many of us in the real world and other coding agents. Now, if we look at how AI has been used in the real world and kind of look at this, I think there's a few different takeaways we could make.

• 3:34Speaker 00

  The first is that generalist models are increasingly being used for real-world problems. And so we're actually seeing general-purpose, like generalist AI models that can do many, many different things, actually be used in the real world, and we see that transition from the left to the right. But I also think that there's a more nuanced observation that we can make from looking at these applications. And in particular, if we look at all of these different applications that are used, or machine learning has actually been useful in the real world and actually been profitable and so forth. But in all of these applications, the customer is making a decision based off of the recommendation of the AI model, more or less.

• 4:18Speaker 00

  And this means that if the customer is ultimately like kind of making the decision, this means that if the system makes a mistake, that's okay, because usually the person can kind of recognize that or decide what to do even despite that mistake. And so even when these sorts of systems aren't perfect, they're still incredibly useful to different people, and there's less pressure on them to be completely perfect. And I think that actually physical AI and robotics is pretty different from this, where if we think about physical AI that are actually operating in the physical world, they have to be directly making decisions that affect the physical world. And this means that they're going to be far more useful when they're operating fully autonomously. And as a result, this requires us to develop physical AI systems

• 5:03Speaker 00

  that make far fewer mistakes than the machine learning systems that have been deployed thus far. Now, one really exciting thing to kind of highlight that has happened recently is a year ago, Waymo passed the quarter of a million weekly autonomous rides, suggesting that it is really possible to develop a machine learning-based system that can operate in a trustworthy and autonomous way directly in the physical world. And I think that brings a lot of hope and optimism for actually doing the same with the rest of AI in the physical world. So if we want to develop general-purpose robots in the real world, I think we need to think about how we're going to make them autonomous for long periods of time

• 5:46Speaker 00

  so that they're actually useful, rather than having them, rather than something where a human is basing decisions on the predictions of the model. So to think about long-term autonomy, I want to ground this in a specific example. And say that we want a robot to make espresso. If we want it to actually be useful for us, we need it to make espresso reliably so that we don't have to babysit the robot very frequently in order for it to help serve drinks. Now, even on its own, this task is really difficult. So actually, operating the porter filter requires very precise and forceful control to insert it appropriately. It also needs to smoothly handle cups with liquid in it and not spill those cups. And it also needs to have an accurate sense of timing,

• 6:31Speaker 00

  which often isn't actually an issue in other areas of machine learning. And not only do we want to do this pretty challenging task, we want to do it with over 90% reliability. So how can we do this? So the first step in machine learning is always to collect some data set, train a model, and evaluate how good your model is. And unfortunately, this rarely works reliably on the very first try. Now, in practice, then, it's a bit better to then iterate on the model that you've developed, where you will try to collect more data or improve the quality of the labels in your data set, or make the labels more detailed, collect more data of the edge cases,

• 7:13Speaker 00

  the scenarios where it's not working well, adjust the balancing of the data set, and so forth. And while this generally improves the reliability of the model, people eventually get tired, and it's hard to get really, really high reliability with a person that's manually tuning this. And so what would be even better is if the AI system itself can iterate on the scenario in which you want it to have high reliability, where it on its own automatically seeks out places where it needs more data, where it needs more supervision. And if we can do this for many, many more iterations because it's automatic rather than a person doing it, then this might be the way to get really, really high, like 99-plus percent reliability from physical AI systems. And so this is the approach that we'll take.

• 7:57Speaker 00

  And this looks a lot like a reinforcement learning algorithm that's trying to attempt the task, learn from its failures, and get better and better on its own. So then, how do we develop a scalable reinforcement learning recipe for robotics? Well, in language models, we have algorithms like PPO and GRPO. And these have scaled to large language models and have enabled really complex reasoning. But there's a bit of a challenge in applying this to robotics, which is that these algorithms have been trained with millions of attempts, or sometimes even tens of millions of attempts, by scaling up the compute, because each attempt is simply running the language model in a data center just by using compute.

• 8:40Speaker 00

  And if we were to translate this very, very approximately to robotics, say we had maybe not millions or tens of millions, but just 1 million trajectories of a one-minute robot task, this is even shorter than the espresso task that I talked about. This would correspond to 700 robot days to get high reliability for that task. Now, maybe this isn't completely out of the question, but this would be quite challenging to do. And that's because the calculus is a little bit different. We're not just running compute to optimize for a use case. We're actually running the robot in the real world and using the hardware and attempting the task in the real world. And so we'd like to have an algorithm that can iterate much more efficiently.

• 9:22Speaker 00

  And there are actually ways that we can make these algorithms a lot more efficient. So there's a couple inefficiencies, like large inefficiencies in these reinforcement learning algorithms for language models. The first is that they spend a lot of time on dead-end trajectories. And maybe this is okay if you're just spending compute on it, but this would cost a lot in the physical world. So we can look at a concrete example. Say that we want a robot to construct cardboard boxes and stack them on the right. And in this trajectory right here, the robot accidentally grabbed two boxes that are flush against each other. And if we let it continue, it would just continue to try to fold that box rather than separate out the two boxes.

• 10:02Speaker 00

  And trying to fold two boxes together isn't useful data that will teach the model how to get better at the task. And so that would be kind of wasting a lot of time on the robot, attempting to go down the wrong path for solving the problem. And so instead of spending a lot of time trying to do that task, what we'll do is we'll actually have a human intervene and show the robot what to do and how to recover from that situation. And so what you can see here is a human is teleoperating and intervening with the robot and showing it that to recover from this situation, it needs to essentially try to separate out the two boxes. It then kind of puts its gripper in, sees if the, then if the robot could autonomously recover. It doesn't autonomously recover, and so the person then intervenes again

• 10:45Speaker 00

  to help it get back on the right track so that we're efficiently using the data on the robot.

• 10:54Speaker 00

  Okay, so this is the first thing that we can do, which is we can show the robot how to recover early, or how to recover, so that we're not spending time on dead-end trajectories, or at the very least, just terminate the episode early. Now, the second thing that we can do is PPO and GRPO and these kinds of algorithms, they make many attempts at a single prompt. And depending on the algorithm, they're essentially trying to estimate for these different responses, what is a good response and what's a bad response. And so even for an individual prompt, they're actually going to kind of roll out like 10 or 50 times for that individual prompt. And they're doing this because they're trying to estimate the value of these different attempts to then kind of upweight or kind of increase the likelihood of good things

• 11:39Speaker 00

  and decrease the likelihood of bad things. But we can actually amortize this cost. Rather than trying to collect a lot of attempts for a single prompt, we can amortize this across different prompts and learn a much more general value estimate of what's good and what's bad, and use this to improve with our autonomous experience.

• 11:56Speaker 00

  And so what this looks like is we can train a general-purpose value function on lots of kind of videos of the robot experience. And this can learn things like if it accidentally unfolds a shirt when it's trying to fold, that that's bad and that's making negative progress. It's shown in red. Or if it's making kind of forward progress, it recognizes that as well. And the same value function can also estimate what's good and bad for a completely different scenario. In this case, for retrieving an item from a fridge.

• 12:27Speaker 00

  And so this sort of kind of general-purpose value model that's kind of predicting basically the time to success can basically significantly reduce the amount of attempts needed to learn how to improve from experience. So with these two kind of improvements to a reinforcement learning system, we kind of have a general improvement algorithm that trains a foundation model on diverse data, then collects experience from that with a human intervening as necessary to help prevent dead-end trajectories, and then train a general-purpose estimate of what's good and bad, the value function, and then use that to improve the model.

• 13:06Speaker 00

  And with this sort of improvement, we're able to kind of fine-tune a foundation model to higher degrees of performance. So in the task of making a latte, in this case, we'll be making a latte in collaboration with a person, where the robot is in charge of making the espresso and the person is in charge of steaming the milk. This is what the kind of model looks like. The model is directly controlling the joints of the robot using the images from the robot's cameras as input.

• 13:37Speaker 00

  And we can see that the model is able to do the pretty challenging task of inserting the portafilter,

• 13:44Speaker 00

  waiting the appropriate amount of time for the espresso to dispense, pouring the steamed milk into the cup. And then the last part of this task is actually the most challenging, where it needs to take a very full cup of latte and transfer that over to the coaster. And so here's actually the observation that the robot sees directly.

• 14:09Speaker 00

  And you can see that the policy is super delicate and able to balance the cup appropriately and smoothly so that the latte doesn't spill.

• 14:20Speaker 00

  So this is kind of gives you a sense of the difficulty of this kind of task. And kind of going back to this reliability question, we took this policy and we ran it not just once, but we ran it for 13 hours straight. And we basically wanted to evaluate, is this policy not only good at making a latte once, but can it do so reliably to the extent that it would be needed to be useful in the real world? And so here's a time lapse of that process. And indeed, we found that the robot was reliable enough to be useful for long stretches of time without making mistakes frequently. Now, the same algorithm isn't specific for making lattes, of course. And so we also applied this to other applications as well.

• 15:05Speaker 00

  Dandelion Chocolate Factory is a few blocks from our office. And so we took a workflow that they typically have a person do, which is to construct these cardboard boxes, label them, and stack them. And we trained our robot to basically do exactly their real workflow and trained it with the reinforcement learning algorithm that I talked about to get a policy that is far more reliable at constructing, labeling, and stacking these boxes.

• 15:31Speaker 00

  And then we also applied this algorithm to folding clothes as well. We wanted, in this case, to not just test how well the model could do one task in one environment, but to do it in many environments. And so these are clothing items that the robot has never seen before in a home that's never seen before, and it's able to do so and act autonomously for an extended period of time.

• 15:53Speaker 00

  Now, videos aren't, don't always kind of show everything. And so we also quantitatively measured the reliability of these models. And we care both about the reliability as well as the speed, like how many boxes can it build per hour? And so we're going to measure throughput, which kind of couples both success rate and speed. And we find that over the phases of training, from pre-training to like an SFT-like stage to an RL post-training-like stage, we see a drastic increase in success rate and speed and throughput. And specifically around a 2X throughput just from the RL stage itself, showing how we can get much greater reliability from reinforcement learning. And for the espresso task, if we look specifically at the success rate,

• 16:35Speaker 00

  we achieved a 90% or over 90% success rate on making espresso.

• 16:42Speaker 00

  So the takeaway for this part is that we can develop a scalable recipe for high reliability of complex robotic manipulation tasks. And we saw, in this case, a 2X higher throughput from using experience and interventions. But most importantly, we saw how we can achieve long-term autonomy in real workflows that people, that people actually care about in the real world. And this is what it's going to take, I think, for robots to be useful in the real world. Now, there's also a lot more work and a lot more opportunities. We actually only ran a few iterations of improvement of this algorithm. And with more iterations, we should be able to see even greater improvement or even greater reliability. And even with this improvement, the robot still makes mistakes. It's also still slower than people.

• 17:26Speaker 00

  And so there's many, like a ton of room for improvement for developing even more powerful recipes. So we've seen long-term autonomy for these different workflows. But there's actually one more ingredient that I'd like to talk about for enabling robots to be autonomous and useful for long periods of time.

• 17:46Speaker 00

  And that ingredient is memory. So you might be surprised to hear that most state-of-the-art foundation models for robotics have no memory or no context. They're just operating on the current sensor observations, the current camera readings, and predicting actions based off of that. And you actually can do short motor skills. You can do repetitive tasks without memory. The videos that I showed before didn't have any context either. But if you want to do a long task that involves multiple different steps in sequence, then memory is critical for tracking progress of the steps that you've completed so far.

• 18:27Speaker 00

  So if it's critical for doing these kinds of long horizon tasks, then why don't these models have any context or memory? There's a couple of reasons for this that are technical, and I'll talk through one of them, which is that if you naively approach memory and try to feed in context, like pass video to a robot foundation model, say that you would just pass in 10 seconds of video. Maybe this 10 seconds of video is sampled at 50 hertz, which is a common control frequency in robotics. And you feed in all four camera streams on the robot, and you use around kind of 256 tokens per image. This corresponds to passing in half a million tokens into your model, which is a lot of tokens. And trying to do that in real time into your model right now is quite challenging.

• 19:11Speaker 00

  Even if you sub-sample to one frame per second, you're still going to be passing in 10,000 tokens into your model, which at least right now is prohibitively expensive for these models. And that's still only 10 seconds of memory.

• 19:24Speaker 00

  So I don't have time to go into the technical details of exactly what we did here, but we also developed a solution for this context problem. And specifically, we developed a system that has memory at multiple time scales. The first is a short-term video memory that has about 10 seconds of video memory, but is done so and computed much more efficiently than naively passing it into the model. And then for longer memory, for memory that spans multiple minutes or multiple hours, we don't necessarily need video of exactly what happened in that past, in that past history. And so instead, we represent memory for those parts in text, where we summarize what happened in text space and then incorporate that much more compressed textual summary

• 20:09Speaker 00

  of what happened over the past 10, 15 minutes into the model as well. And with this sort of kind of memory at multiple different time scales, we're able to enable robots to do tasks that can operate for 10 or 15 minutes at a time completely autonomously. And what's different from the previous slide or the kind of what I showed previously is that this task isn't repetitive. So this is going to be a 10 to 15-minute task that's involved cleaning a kitchen. And the robot isn't just repeatedly making espresso over and over again. What it involves is wiping the counter with a sponge, kind of then drying the counter with a paper towel, throwing away the paper towel. Next, it's going to put away the mustard into the fridge.

• 20:54Speaker 00

  Then it will put the dishes away into the, into the cabinet, wash some of the dirty dishes in the sink, and so on. And so by incorporating memory, it's able to do a task that requires keeping track of all of these different steps that is done to clean the kitchen and successfully operate for 10 to 15 minutes completely autonomously.

• 21:18Speaker 00

  Great, so those were a couple ingredients for long-term autonomy. Now I would like to build on that and actually take those ingredients and put it into a general-purpose model that can do everything that I showed before, but also can do that in a single model and can do some other things as well.

• 21:37Speaker 00

  And to think about developing such a general-purpose model, I think it's really helpful to contextualize like where robotics is at within the timeline of other developments in generalist AI. And if we think about how kind of generalist AI systems have evolved over the past like 15 years, I think the first major milestone was in 2012, when we saw that a deep learning system trained from scratch, this was the first time that it topped an external benchmark and all of the previous methods for that benchmark were specifically designed for that application. So all of the previous methods, specifically this was the ImageNet benchmark, were designed specifically for image classification,

• 22:20Speaker 00

  whereas this was the first time that a deep learning-based system actually outperformed those more specialist systems. And this is kind of a much more general algorithm that wasn't specifically designed for image recognition. And then just a couple of years later, we found that we weren't just training algorithms from scratch, but we were able to get models, like pre-trained models, that are useful for fine-tuning to downstream tasks. And so it became the norm to take a model that was pre-trained on ImageNet and then fine-tune it on a downstream task. And we actually saw better performance from using that pre-trained model like BERT or like an ImageNet pre-trained model.

• 22:56Speaker 00

  And then from there, the, I think the next big phase and the next big transition in generalist AI models wasn't using pre-trained models, but kind of moving from a pre-training fine-tuning regime to a regime where we're just using generalist models out of the box. This was with models like the start of like GPT-2. And then, of course, all, like, almost all the models that we interact with today worked just out of the box without, without fine-tuning. Or at least most of the consumer models. There's actually other, other models that still use a lot of fine-tuning. And then I think one other kind of milestone that I want to highlight was a milestone in 2021, where we think I saw the first signs of compositional generalization in these models. And kind of one specific instance of that was with DALL-E,

• 23:40Speaker 00

  and I'll talk a little bit more about that in a later slide. So this is how kind of generalist AI has advanced over the past 15 years. And then meanwhile, if we kind of think about physical AI, even just three years ago, in 2023, it was extremely common for people working on robotics to collect a bespoke dataset from scratch for an individual project and train from scratch on that dataset. And this is analogous to, say, collecting ImageNet from scratch and training on ImageNet, or training on the dataset that you just collected from scratch on that dataset. And so if you want to develop a general-purpose model, if you have to collect the dataset from scratch for every single project,

• 24:25Speaker 00

  you're probably not going to make a lot of progress. And so until just a few years ago, I think we are kind of pretty far on the left of this timeline. And until recently, I think that, more recently, I think that we've been kind of in the 2014 phase, where we have some good pre-trained models, but we haven't really been truly in the regime on the right.

• 24:47Speaker 00

  So how do we get to that regime on the right? Specifically, how do we develop a single general-purpose model that works out of the box and also shows compositional generalization?

• 24:58Speaker 00

  So this has two goals. The first is out-of-the-box model. This is analogous to going from like BERT to GPT. And right now, the best robot performance, if you want to get like your model to perform the best that it can on a given task, it always requires fine-tuning. So some of the videos that I showed at the beginning, they were fine-tuned models to do things like unlocking a lock. Other work that we've done on measuring human-to-robot transfer also needed fine-tuning to get the best performance. And of course, all of the videos that I showed with RL post-training were also fine-tuning on an individual task to get the best performance on something like making espresso. But if you have to fine-tune a model, you actually aren't getting a general-purpose model for the things that you want it to do,

• 25:42Speaker 00

  because you have to fine-tune it for each individual thing. And so our first goal is to move towards a single general-purpose model that can actually do all of the things that you want it to do. And then the second goal that I mentioned is compositional generalization. And this is inspired by the DALL-E result from 2021. And I think that this was a really important and exciting milestone because of the compositional generalization that it achieved. And specifically, when you have compositional generalization, when you can basically bridge the concept like of an avocado and a chair and show that you kind of can combine those two, it means that the model has at least some kind of conceptual understanding of what an avocado is and what a chair is, to the point that it can combine them into something that exhibits both concepts at the same time.

• 26:29Speaker 00

  And then second, it means that you have some degree of data efficiency, where your data doesn't need to cover all of the possible combinations of concepts represented in your data. You don't need pictures of avocado chairs in your dataset in order to generate something like this. Or you don't need combinations of other things that you might ask the model to do when it's deployed. Now, even back in 2021, it wasn't perfect, but these signs of compositional generalization were really exciting for demonstrating these two attributes of the model.

• 27:01Speaker 00

  So we have these two goals that we'd like to do. We like an out-of-the-box model and compositional generalization. Now, the tried-and-tested recipe for developing this kind of model is to first, take a sufficiently large and diverse dataset, and to second, train a model with sufficient capacity. And so what we're going to do is we're going to do that. We're going to try to use all of the data that we have available. This includes really diverse robot demonstration data, including really low-quality demonstration data. It's also going to include policy rollout data, basically attempts from the robot of doing the task. Basically, all of the training data that was used for reinforcement learning for the previous tasks will be included in the training recipe. We're also going to include videos of humans,

• 27:47Speaker 00

  and we're also going to include data from the web. Basically, all of the data that we have. And then to train a model with sufficient capacity, of course, we'll train a model that's large enough. But to fit data that's so heterogeneous, we also find it particularly important to prompt the model with all of the context that it needs in order to predict actions. And we found that this idea was really the key unlock to using this kind of data and this data of this kind of degree of heterogeneity. And specifically, what this looks like is we're going to train a foundation model that takes as input the memory like that I mentioned before, an instruction of what to do.

• 28:31Speaker 00

  But it's also going to take as input a subtask instruction of kind of what the next immediate thing it should do is. It'll also take as input metadata that kind of indicates the quality of the data, the length of the episodes and so forth. And this metadata gives it a lot more information about how it should predict the next action. And then optionally, we'll also train the model with a sub-goal image as a prompt to the model. It's essentially saying, like, a few seconds from now, you should try to reach something that looks like this image.

• 29:01Speaker 00

  With this detailed prompting, we find that the model can really make use of much more heterogeneous data. And I'll show some comparisons later that really kind of show how important it is. And now to actually deploy this model, we need to then provide things like this subtask instruction and sub-goal images. And so with that, we can train a high-level policy that predicts the subtask instruction, like what to do next, what is the next like subtask for the task of cleaning the kitchen. And we'll additionally train a world model to generate images for what the robot should do next as sub-goal image conditioning.

• 29:36Speaker 00

  And so with this, we'll train the single model on, with those attributes, on all of the diverse data that we had available. And here are some examples of what that single model can do. So all of these videos are from a single model, specifically a model that we called the Pyo7 model. And on left, you can see it doing things like folding a collared shirt. On the top right, it's doing a really precise assembly step, where it needs to insert a screw and drill that screw into a robot arm. And on the bottom right, the robot is replacing a trash bag in a trash can.

• 30:15Speaker 00

  And so we had two goals at the start of this. The first was to move towards an out of the box model. And even those videos showed that out of the box, the model is able to do quite a bit. But really, the key question here is how does this pre-trained model compare to the specialists that were trained specifically for coffee making, specifically for box building that I talked about previously? And if we measure the throughput and the success rate of this single Pyo-7 model versus the fine-tuned Pyo-6 model, we see that the, across the board, the single Pyo, like, pre-trained Pyo-7 model matches or outperforms the fine-tuned specialists that were developed with reinforcement learning post-training for those downstream tasks.

• 30:59Speaker 00

  And so we see that it's able to match the performance. The specialist is also holds for SFT specialists, not just RL post-trained models as well, suggesting that we do indeed have a single model that can do a lot of different tasks with a really high degree of performance out of the box.

• 31:16Speaker 00

  Okay, so that was the first goal of out of the box models. And the second goal is compositional generalization. And there's a few different ways to measure this. There's many different ways you might try to combine concepts in robotics. The first test that we wanted to do was to see if a robot could interact with an appliance that's quite rare, like an air fryer. And this is kind of an example. We basically wanted to see if it could open an air fryer, put an sweet potato in the air fryer, and close the air fryer. And we picked this because we thought that the dataset didn't have any air fryers in it. We didn't intentionally collect any training data with air fryers. After we did some analysis on the dataset, we actually found that our dataset was so diverse that it did actually have three episodes with air fryers in it.

• 32:01Speaker 00

  We expect that they likely weren't having an impact on, and that even if you didn't include those exact three episodes, this likely would still work. But what we found generally is that the, the robot was able to interact with appliance that was basically hardly represented at all in the training dataset and combine the skill of interacting with it, like opening it, closing it, and so forth, with this object that it hasn't seen before.

• 32:28Speaker 00

  And then after instructing it, like Lucy did, we can train a high-level policy to do this task fully autonomously, and you can see the robot doing that in this video. So that's the first form of compositional generalization. The second compositional generalization test that we wanted to look at is we wanted to see if we could compositionally generalize between tasks and robot platforms. And so we wanted to take a robot platform called a Byarm UR5E robot. It's actually a very large, kind of industrial robot platform. And we wanted to see if it could fold clothes, despite the fact that we didn't collect any folding data on this robot platform. So specifically, we had data of folding clothes, like folding a shirt, on the robot platform that's pictured here on the left.

• 33:14Speaker 00

  And then we wanted to see, like, out of the box, without collecting any folding data on this very different robot platform on the right, could the robot successfully do the task? And what we see in this video is the, we indeed did see that it kind of compositionally generalized in this manner. The first time we saw the robot do this, we were like, like floored, because we, like, there was no training data for this, for this task. And the robot here is quite different from the other robot, not just in size, but also in the lengths of the linkages of the robot, in the configuration of the joints of the robot, and so forth.

• 33:52Speaker 00

  Okay, we can, and the, this is a 1x speed video, so it's not the, the fastest thing. And obviously, if you haven't seen any training data on something, you might not, if it's literally the robot's first time folding a shirt, it might take a few attempts. But eventually, it will get to the folded shirt. You can also see the generated sub-goal images on the top left. So those, those are basically the models trying to generate images that will make progress on the folding task. And then those are passes input to the model, and we see the folded shirt here. I think it's going to make a couple small corrections at the end to try to make it a little bit smoother.

• 34:30Speaker 00

  Cool, so the takeaway here is that both in terms of like, language, object interactions, and in terms of task, robot interactions, we see strong signs of compositional generalization in this model.

• 34:42Speaker 00

  Okay, and then quantitatively, we also see that as we kind of get to these more advanced models like the Pyo-7 model, the performance of folding towels and folding shirts on this platform that hasn't seen before increases dramatically. And it even approaches the performance of human teleop despite the fact that we didn't have any robot-specific training data for folding clothes.

• 35:09Speaker 00

  And then for the last experiment that we did here, I think this may be perhaps the most interesting experiment, is we wanted to test, how important are the two ingredients that I mentioned? How important is diverse data? And how important is this sort of like capacity or detailed prompting for the kinds of results that I showed? And so if we remove the most diverse data from the model training, shown in like the grayish color, we find that the performance on held-out tasks decreases dramatically. Whereas if we just take out a random 20% of the data that's less diverse than the most diverse subset, the performance only decreases a little bit. So this suggests that actually having really diverse data plays an important role in enabling it to generalize to new tasks.

• 35:55Speaker 00

  And then we tried to also oblate the fact that we are prompting the model with metadata. And for this experiment, we looked at with and without prompting with metadata. So with prompting is shown in yellow, and without prompting is shown in the gray color. And with prompting, it helps kind of significantly. But the most interesting thing is if you look at, when you add, so this plot is showing as you add more and more data, and specifically as you add more and more low-quality data, what is the performance? And without metadata prompting, when you add lower-quality data, from 80% data to 100% data, the performance actually decreases, which is perhaps not too surprising because you're adding low-quality data to your data mixture. Whereas with the metadata prompting, the performance actually increases when you add

• 36:40Speaker 00

  that low-quality data, suggesting that it's actually able to make, get a lot more juice out of even low-quality data when you include this kind of prompting.

• 36:51Speaker 00

  Cool, so the takeaways here are that we found that we're able to train a single model to control the robots that matches or exceeds the performance of specialized post-trained models, kind of like going from a BERT-like pre-trained model to a model that really works out of the box, like GPT. We also saw strong signs of compositional generalization in a DALL-E-like way. For example, in compositionally generalizing skills applied to appliances and skills applied to new robots in ways that weren't seen in the training data.

• 37:23Speaker 00

  Cool, and then all the like videos and experiments that I showed were just evaluating the model out of the box without any post-training. And the paper and the technical report online have a lot more experiments and a lot more details. Okay, so we talked about long-term autonomy. We then showed how we can kind of develop that in a single general-purpose model.

• 37:45Speaker 00

  Where are we at now?

• 37:48Speaker 00

  The first thing that I'll mention is if we go back to the timeline of generalist AI, I think that we now kind of firmly have physical intelligence in the right side of this timeline. We're kind of firmly more in like a GPT and DALL-E-like era for robotics and physical intelligence, which is really exciting. And I think that, and we've kind of went there in, in just a few years. And lastly, we also have these models that are actually deployed in real-world circumstances. So the two videos on the top are actually two YC companies, Ultra and Weave, that have taken Pye models and post-trained them to do, in deployment, to do tasks like folding laundry and, and packaging in a warehouse.

• 38:31Speaker 00

  The kind of video on the bottom, bottom left is the video that I showed previously. And this kind of model works for a really diverse set of robot embodiments. The ones on, on kind of the top and the left are kind of a more standard by manual platform. But it also can be adapted to things like drones, quadcopters, surgical robots, and on the bottom right, for things like tractors. And so this is really kind of truly showing how physical intelligence can make an impact, not just in demos and research and so forth, but actually in real-world deployment. And I think that we'll, as we go, we'll start to see robots that are, more and more robots actually deployed for real in the physical world with all the advances that we've been

• 39:15Speaker 00

  seeing over the past few years. Awesome, so the last thing that I'll mention shamelessly is that we are hiring at physical intelligence. So if you're excited about some of the stuff that I talked about, we encourage you to, to take a look at some of the open roles and apply. And yeah, definitely have time for, for questions and happy to, happy to get all your thoughts. Thanks.

• 39:44Speaker 00

  Okay, so the first question is how far away are we from a chat GPT moment for robotics? And what will that look like? So I'll start with the second part, actually, which is that I I'm not sure it will really look like the chat GPT moment that we saw in language models, which is that with something like chat GPT, we saw like what like it passed like what a million users in five days. I think that the distribution channel for physical models is going to be slower, unfortunately, because you actually need a physical robot there. I mean, I think that we've seen for something like Waymo, it's like the rollout. It's actually been like incredible to see the rollout that they've had, but it still takes time to actually deploy things on physical devices.

• 40:26Speaker 00

  So I don't know if we'll have a single moment that has the distribution that chat GPT had. At the same time, in terms of the capabilities of these models, I think that we are really starting to get to the point where these models are actually useful in the real world. And I think that getting to the kind of the capabilities of chat GPT, I I think is yeah, very much on the horizon in the next few years.

• 40:49Speaker 00

  Cool, the second question is when should a small team switch from scaling per site models to a generalist policy? And what does that transition actually look like? What signals tell you it's time? So this is a good question. I think that the I mean, at the very least, I actually think that just starting with a generalist policy and then fine tuning it, even like right off the bat, can be really effective. And we've and like fortunately, a lot of generalist policies, like they're actually really powerful generalist policies that are open source. And so the PI zero and PI O five models are are open source, for example. And we've seen a lot of people get a lot of use out of those models already. And we're also working with a lot of partners like the tractor company, like Ultra, like

• 41:34Speaker 00

  we've to actually develop to take our most recent models and and get even more juice out of even more powerful models for their own applications. So even right off the bat, I think that you can use them. The only scenario in which I actually wouldn't use them is if if you're really in a really constrained environment. I've I've talked to some folks working on surgical robots that are like in an operating room in the basement with no internet connection and a really bad GPU. And so sometimes like it's just really hard to use a larger model. But you still can do local inference with on a workstation with these models. And so I think that, yeah, right away, I just taking like PI O five or or your favorite model and and fine tuning it is is the way to go. And I think that we'll see lots of these small companies.

• 42:18Speaker 00

  And I think there's like, yeah, so much work to do in terms of actually getting these robots to work with this technology in the real world.

• 42:25Speaker 00

  Great, the next question is, given how fast robotics is moving in industry, what are the real advantages and drawbacks of doing a PhD today? Especially for someone who wants to go into industry afterwards? So I was not planning to do a PhD. I was always planning to go straight to industry. My parents are engineers and worked in industry. And I kind of thought that the way to have impact was to go to a company and so forth. And the I mean, my dad even told me that he wouldn't hire someone with a PhD. And so I was like, maybe I shouldn't get a PhD if I wouldn't be able to get a job. But there's also he's in a different field as well in civil engineering. And so

• 43:09Speaker 00

  the but at the same time, I I think that like a PhD is an incredible opportunity. And I I I love my PhD. And like, obviously, it depends a lot on the advisor, depends a lot on like what you would be doing and so forth. But I think a PhD is an incredible opportunity to first learn a lot about how to handle uncertainty, how to pick good problems to work on. And I think that basically, like in in research, you don't no one kind of even gives you the problem to work on. You have to pick the problem. You don't know for the problem that you picked, if it is achievable to make progress on that problem in a six month time span, two year time span, 10 year time span. And so you learn about how to deal with that uncertainty. And then the so that's really useful. And then it also

• 43:54Speaker 00

  it also is an opportunity to like do amazing research and do whatever. In many cases, do like have a lot of freedom to work on the research that you find most exciting. And so I think that today is still, yeah, an amazing opportunity to do to work, to learn about uncertainty. And I think that learning about uncertainty is really useful in the startup environments, in being at the frontier of AI, because we don't know now, like no one knows what the best route is to make these models more and more powerful. At the same time, there's also a lot of incredible opportunities in industry. I think that the first in terms of what goes into developing like everything that I showed, it's not just the research. There is a whole like software stack that needs to run on the robot, needs to run reliably. There's obviously everything on the hardware side.

• 44:38Speaker 00

  And then there's also even the machine learning infrastructure, the data infrastructure and all that. And so even with like, you don't need a PhD necessarily to do a lot of that engineering work. And then on the research side, there's often opportunities to get involved as well. And a lot of research is engineering these days as well. And so I think it kind of depends on, yeah, it's a very personal decision and what you want. I I even today, I think I probably like kind of retrospectively would want to do a PhD just to learn about how to handle uncertainty, to do research, because I really love being at the frontier and thinking about challenging problems. But there's also a lot of really amazing opportunities in in both paths.

• 45:25Speaker 00

  Okay, the next question is, large language models learn from the internet, but robots don't really have an internet scale data set, a physical experience. What's the robotics equivalent and how do we get it? So in robotics, well, so maybe in language models to start off, the data on the web is language data. And not all of it's high quality, but some of it is like really informative and useful. And it is data that is in kind of the that that kind of reflects a lot of what the like what you want a model to do. Like you want it to be able to predict text and be able to complete text, answer questions and so forth. And there's a lot of questions being answered on the internet and a lot of text that's being completed on the internet.

• 46:08Speaker 00

  And so in general, with machine learning, you want train to match tests, test. And so you want the thing that you're going to be training your model on to be reflective of the thing that you want, you're going to be asking it to do later on. And so I think the equivalent in robotics is data of robots operating in real world circumstances. And the way that we approach it at physical intelligence is to collect data, like to collect robot experience of robots doing all sorts of tasks. You can collect this with teleoperation as like kind of to get initial data of robots doing useful things. But in the long run, I think it will also contain a lot of autonomous experience of robots deployed attempting things. And just like how we see in language models, how now a lot of time is spent actually generating

• 46:52Speaker 00

  data, generating synthetic data by actually running the model and having it think through things. I think a lot of the data in the future in robotics is going to be the robot attempting to do lots of tasks in lots of real world circumstances. And so, yeah, I think that that's kind of what it looks like. I also think that there are other possible sources of information that's really useful for model training, like videos of people doing things like YouTube, like web data and captioned images that tell you like this is a kitchen that has a fridge on the right of the sink and so forth. And all of that data, I think, can be really useful for developing a kind of frontier multimodal model that can control robots to do things, reason through how to do a long task and also control the robot to do those tasks.

• 47:37Speaker 00

  I think that there's no substitute for the robot experience itself. You can't just like, if you watch a human do something, like if I watch Roger Federer play tennis, it doesn't mean I can't I can play tennis as well as him, unfortunately. And likewise, robots can't like watch a person doing something and then figure out how to do it themselves directly. They really need their experience on their own platform to learn effectively. And I think that we will need large data sets. I think that doesn't mean the human video isn't useful. It's useful to watch Roger Federer play tennis. But the but the actual experience on robot platforms will be a critical component of developing an analogous data set for robotics.

• 48:16Speaker 00

  The next question is, is it possible that general purpose robotics models get democratized via open source the way that large language models did? Or will the cost of embodied data and hardware keep the best models concentrated in a few well-resourced labs? So I think this is a good question.

• 48:34Speaker 00

  So I do think the cost of embodied data and hardware will like could very much make this look different, because I think that it's harder to get data, even to like distill a model, for example,

• 48:48Speaker 00

  just like readily on the internet. I also think that we've seen a kind of pretty large data sets get open sourced as well and pretty powerful models get open sourced. The I think it's really hard to say exactly what will happen.

• 49:04Speaker 00

  And so, yeah, I I don't know.

• 49:08Speaker 00

  The one thing that I will say is that I like with language models, even aside from like democratization like Kimi and and like really getting models that like perform at the state of the art, even then, companies that are focusing a lot on closed source models are also doing a lot of open sourcing. And so there are like, like, like Juma, for example, and the GPT open source and so forth. I think these these companies like to support open source because it actually helps build the ecosystem around the things that they're building. And so I imagine there being, I can be optimistic of that there will be a strong open source community regardless, but I don't know if it will exactly play out exactly the way that language models play out.

• 49:48Speaker 00

  Okay, the next question is, does the model output raw motor commands directly? Or does it output a target hand position and let a controller solve for the joint angles? And what makes that the right level to learn at? So, all the models that I showed were outputting target joint positions. So like, what is the angle of this joint? What is the angle of this joint and so forth? That you want to hit? And then there's a controller, like a PD controller, that is trying to then hit that target position for those joints.

• 50:17Speaker 00

  The model actually is also trained to predict target gripper, like, yeah, gripper positions. Like, where in 3D space should my gripper be? And you could also use that as well and back out the joint positions. You could also, like, another thing you could do is you could go directly to like motor torques, or to like voltages or efforts.

• 50:37Speaker 00

  There are pros and cons of different options. We have found controlling joints and controlling in like the 3D space of the gripper to be, to both work well.

• 50:51Speaker 00

  And so, yeah, there's pros and cons. I think the one thing that would be nice about going directly to the, like, the voltages is that you could also get a more stiff kind of output, or a less stiff output. Whereas is the controller, if you have a fixed controller, then you, you're not letting your model control that, that aspect. And so, yeah, there's different pros and cons.

• 51:16Speaker 00

  Yeah, I think that the, it's like, what we're working seems to work. It doesn't seem to be a bottleneck. And I often like to focus on the things that seem to be bottlenecks versus, versus things that don't seem to be bottlenecks.

• 51:30Speaker 00

  Okay, next question is, do robots need something like imagination? The ability to picture what should happen next before they can become truly useful? So, the Pyo7 model that I showed has something like this, where it can kind of imagine what a future image should look like and then try to accomplish that.

• 51:51Speaker 00

  We found that that leads to improvement. And we saw on the shirt folding example, we saw like a quantitative bump from using that sort of imagination compared to not using it. At the same time, I think that the model actually performed surprisingly well without that as well. And we were actually hoping, we were thinking about writing an entire paper, an entire technical report just about that capability in that model. But the, the model without that was so good that we felt like we needed to actually have that play a bigger part of the story. Cause it, it, that seemed like that was really delivering in terms of actually getting really strong results. And so, I mean, it seems like one design choice. I think that the, uh, and I, I don't, I think it's hard to say if it's going to be a critical component or not.

• 52:33Speaker 00

  The, I guess the good news with, with these kinds of models is that the, if you develop a good dataset, you can kind of run experiments and you can continue to test things with the dataset that you have, uh, quite effectively. I also think that being able to predict the future seems like a very relevant objective compared to predicting future actions. And the, so that should, I would imagine, kind of help in terms of learning from all the data that you have available to you. So, yeah, hard to say if it'll necessarily be a critical component or not. It seems like empirically, so far, it seems to help. Although perhaps not as much as you might expect. And even without that imagination, the robot can do pretty incredible things.

• 53:14Speaker 00

  Okay, um, next is, right now, it seems that robots are doing amazing tasks, but in a very slow manner. What is needed to improve the speed? I'm really excited about improving the speed. The, and we, we did see speed improvements from reinforcement learning. We also have another release called the RL token, that where we showed actually even faster speed and actually faster speed than human teleop. I think one of the bottlenecks is that when you teleoperate robots to do things, which is the easiest way to teach a robot to do something, people are kind of slow at teleoperating the robot. We have a couple of projects in the pipeline that I think have really promising results in terms of getting fast policies. And so, I think more to come there.

• 53:56Speaker 00

  And yeah, I think it's, either you need to figure out how to make the data faster, or you need to figure out how to be faster than the data. We see the evidence of being able to be a little bit faster than the data. And in terms of the next steps, it's either, either to kind of get, go even further than that, or, or make the data faster.

• 54:16Speaker 00

  Cool. What's the most surprising task you've seen a robot complete recently? What do you want to see it do next?

• 54:23Speaker 00

  So, the most surprising thing was not really a task, but when we were working on Pyo7, I was, I personally trained one of the policies for, like, some of the initial tests for assembling this pinwheel. And when I was working on training, training it to construct the pinwheel, one thing that really surprised me was, in all the data, we kind of carefully controlled, controlled the strategy for how to assemble the pinwheel, where you basically will take the, the kind of pre-cut piece of paper and take a little pin and insert the pin into a hole in the paper. And in all of the data, we picked up the pin with the right hand and picked up the paper with the left hand and inserted it. And the robot kind of started doing that. And then it actually made a mistake.

• 55:05Speaker 00

  And the paper ended up on the right side and the pin ended up on the left side. And what the robot did is it picked up the paper and it picked up the pin with its left gripper. And it put the pin with its left gripper and inserted it into the paper with its right. And it had never seen data of inserting the pin with its left gripper. And it kind of showed that even that wasn't in the post-training data at all. Like, it was only, like, and it wasn't even in pre-training either. The robot essentially had learned this sort of equivariance between its left hand and its right hand, so that it could actually transfer behaviors from one hand to another, despite the fact that that was never in the data. And so, that was like a really cool moment. I feel like, I don't know if other people were as excited about it as I was when I shared it with some people.

• 55:47Speaker 00

  But it kind of shows this kind of emergent capability in these models that I hadn't seen before.

• 55:54Speaker 00

  And then, yeah, in terms of what I'd love to see, I don't know.

• 56:00Speaker 00

  I, I love seeing robots do anything. And the, I guess, I think that there's still a lot of, a long way to push in terms of reliability for robots being able to do tasks for really long periods of time. I don't necessarily think that much about individual tasks, but more so about capabilities and how to get the next capability from these models. The, yeah, and so, I, yeah, anything. I guess the, yeah, a robot doing anything always gets me excited, even if it's something that hasn't been done before. I guess one thing that we've been doing recently is having robots use knives to slice vegetables. I think that there's a lot that you can do there, once you actually can use knife safely, which is one thing that we've done recently.

• 56:43Speaker 00

  Okay, and then the last question is, how can someone break into robotics from a software engineering background? Great, um, so I think that the, I guess first, there's a lot of software engineering in robotics. And so, there's, I think you could try joining a robotics company as a software engineer. And then another thing that I would mention, and I've actually seen someone take this path. So, someone who actually now works at Physical Intelligence, her name is Jenny. She worked in trading, algorithmic trading for a while. Then she worked at Harvey and was doing legal stuff. And she was really excited about robots. And so, she bought a cheap robot and basically in her bedroom,

• 57:27Speaker 00

  like, played around with it and tried like fine tuning an open source model and trying to get it to do something. And then she shared what she had done. And she sent me a cold email and was like, hey, could I, like, I'm interested in working in your lab. And it seemed like her profile was promising and that she actually like, actually got out there and like tried it and done it. And that she was really excited about that. And so, the, and now, now she works at Physical Intelligence. The, yeah, I think just like getting your, getting your feet wet, trying stuff out and learning from that experience. And then, yeah, using that experience to kind of, yeah, share with people, have it on your resume and so forth, I think is a great way to do stuff. And fortunately, there's a lot of open source stuff out there that can allow you to get started on those kinds of things.

• 58:12Speaker 00

  Great, that was the last question. Thanks, everyone, for listening.