Teaching Atlas how to learn

The Toyota Research Institute has been doing some incredible work teaching robots to rapidly learn and perform tasks autonomously – now, it’s bringing its Large Behavior Model tech to the extraordinary Atlas humanoid in partnership with Boston Dynamics.

Humanoid robot hardware, believe it or not, is probably good enough already. More than a decade’s worth of work at Boston Dynamics has resulted not only in an incredibly athletic and capable hydraulic Atlas robot, but a slew of emerging commercial competitors from Tesla, Figure, Agility, Sanctuary, Fourier and many others. These remarkable robot bodies will continue to improve, but they’re already good enough to perform all manner of useful work.

The software is the problem. If you need a coding team to teach a robot a new behavior, it’s scarcely better than today’s conventional production robots. But developing a general-purpose humanoid robot that understands the world and how to interact with it in flexible and adaptable ways is an enormous task.

The answer is AI, of course, like the answer will soon be for everything – but AI needs to be trained on lots of data. ChatGPT, Grok, Llama and Claude all benefit from the insane quantity of (largely written) data humanity has accumulated over the centuries. Large Language Models (LLMs) have advanced so far so quickly, because language is a highly compressed representation of reality, crunched down to such small file sizes that vast amounts of it can be processed.

There’s much less data available to help robots learn the basics of movement – other than video, which doesn’t tell the whole story about why somebody made what movement. And they really do need to learn things from the ground up. Hence the idea of a ‘Large Behavior Model,’ or LBM – a way for robots to slowly build up the basic movements they can use to interact with the world, and combine them into more complex movements in service of a task or goal, in a similar way to how LLMs have developed an ‘understanding’ of human language and learned to interact with us.

If you haven’t seen the LBM work the Toyota Research Institute (TRI) was doing last year, take a moment to catch up:

In essence, The TRI team developed a telepresence system allowing human pilots to ‘drive’ robotic arms, looking through VR goggles fed from the robot’s cameras to see exactly what the robot sees, and equipped with haptic gloves to let them also feel what the robot’s tactile sensors can feel.

Then, with the human pilot limited to the exact set of ‘senses’ that the robot has, they set about doing a bunch of tasks, a lot of them in a kitchen setting. They’d spend a couple of hours doing the same task over and over from different starting points, correcting their mistakes if they made them, and marking each attempt as a success or failure.

From there, the robots would spend some time ‘thinking’ about the problem, effectively running millions of different simulations of the task while adding random variables and starting points, grading their own performance each time according to their own understanding of success and failure modes.

And it worked. The TRI team had taught its robot arms more than 60 complex behaviors by September last year when we first saw the video above. Researchers reported that they’d often spend an afternoon doing the piloted training, then go home as the behavior learning system ran its simulations overnight, then they’d turn up in the morning to find that the robots were now able to do the task by themselves, and in a fairly flexible manner.

It was remarkable stuff, and we’re fascinated to learn how far it’s come in the last 12 months, given how astonishingly quickly things are progressing across all fields of AI. But it was also fairly limited research, done using pairs of robot arms rather than whole bodies.

Well, that’s about to change. Boston Dynamics is the absolute gold standard in robotics research, and has been for decades. The old, hydraulic Atlas humanoid will go down as one of the most groundbreaking and significant machines in the history of robotics.

And the fully-electric new Atlas, which you may have seen on New Atlas, made its public debut just five months ago. This remarkable evolution lost some of the explosive power that made the original Atlas such an extraordinary gymnast – but made up for it with fully swivel-capable joints all over its body, allowing free rotation at the hips, shoulders, waist, neck, biceps and thighs, so any given section of its body can face any direction. From gymnast to contortionist … Check it out:

It’s a truly remarkable looking robot, already scratched, dented and looking very second-hand in typical Boston Dynamics fashion, but we’ve seen curiously little of the new Atlas in the last five months to determine exactly where it’s at and what its current capabilities are. Well, other than this: we know it can rip out a set of push-ups.

So today’s news is very exciting; the undisputed grandmasters of humanoid robot hardware, teaming up with a leading team in AI LBM development in order to advance the useful capabilities of humanoid robots.

“There has never been a more exciting time for the robotics industry, and we look forward to working with TRI to accelerate the development of general-purpose humanoids,” says Robert Playter, CEO of Boston Dynamics, in a press release. “This partnership is an example of two companies with a strong research-and-development foundation coming together to work on many complex challenges and build useful robots that solve real-world problems.”

“Recent advances in AI and machine learning hold tremendous potential for advancing physical intelligence,” adds Gill Pratt, chief scientist for Toyota and CEO of TRI. “The opportunity to implement TRI’s state-of-the-art AI technology on Boston Dynamics’ hardware is game-changing for each of our organizations as we work to amplify people and improve quality of life.”

The partnership aims to rapidly develop whole-body behavior models for the Atlas robot, but also for other humanoid platforms that TRI may go on to work with. It’ll be interesting to see what different kinds of telepresence training hardware are brought to bear on the problem, since Atlas is so much more complex than the simple bimanual setups TRI was working with originally.

Boston Dynamics

Ultimately, though, it’s still unclear whether Boston intends to scale Atlas into a commercial product. And scale may be crucial here; companies like Tesla and Figure are designing their humanoids with mass manufacturing in mind, aiming to deploy hundreds, then thousands of them out in the world doing small, simple, useful tasks. There, they’ll see an extraordinary range of things happening, collect great scads of real-world data, and use that data to drive swarm-based learning.

That’s the approach Tesla claims makes it a world leader in autonomous cars; there’s millions of these things on the road already, constantly watching and contributing to the knowledge of the whole. AI is a big-data game, and whoever gathers the most data and uses it most efficiently will win, according to this model. And the prize, according to people like Elon Musk, is possibly the biggest product in history, a transformational labor-replacing machine that could eventually take over basically any physical job.

While Boston Dynamics has been miles ahead of everyone on humanoids for at least a decade, Atlas has been specifically designated a research platform. The company has restricted its commercial activities to smaller, practical quadrupeds like its Spot platform, and its single-arm, heavy-lift Stretch box handler.

Perhaps it’s telling that this pioneering company doesn’t seem to think humanoids are ready to get to work just yet, and will need a few more years in the lab, painstakingly putting together the building blocks of physical behavior.

The promise of general-purpose humanoids is so massive, and the challenge so great, that there’s sure to be plenty of surprises on the path. This does feel like a field where we’re watching the history of the future happen in real time.

Source: Boston Dynamics