Boston Dynamics is taking a new step with Atlas, its electric humanoid robot, by teaming with Toyota Research Institute to advance AI-based robotic intelligence. The collaboration is aimed at a difficult goal in robotics: building machines that can learn more broadly, act more autonomously, and eventually handle a wider range of real-world tasks.
Why This Atlas Partnership Matters
The announcement brings together two organizations with different strengths. Boston Dynamics is best known for advanced robot hardware, including Spot and the new electric Atlas. Toyota Research Institute, or TRI, has been working on robot learning, including large behavior models, often shortened to LBMs.
Those large behavior models are described as operating along similar lines as large language models, the technology behind platforms like ChatGPT. But the target is different. Instead of producing language, the work is focused on helping robots learn behaviors that can be applied in the physical world.
That distinction is central to the challenge. A chatbot can train and operate in a digital environment, while a robot has to interact with objects, rooms, surfaces, tools, and its own mechanical limits. When learning involves physical action, every failed attempt can take time, create wear, or even risk damaging the machine.
TRI’s Robot Learning Work
TRI’s work has already shown why the partnership is notable. At last year’s Disrupt conference, institute head Gill Pratt described research in which the lab trained robots to perform household tasks such as flipping pancakes. According to Pratt, the lab reached 90% accuracy through overnight training.
Pratt explained that machine learning has often required very large amounts of training data. For physical robotics, that creates a practical problem because a machine cannot endlessly repeat actions without time and durability becoming constraints.
“In machine learning, up until quite recently there was a tradeoff, where it works, but you need millions of training cases,” Pratt explained at the time. “When you’re doing physical things, you don’t have time for that many, and the machine will break down before you get to 10,000. Now it seems that we need dozens. The reason for the dozens is that we need to have some diversity in the training cases. But in some cases, it’s less.”
That shift, from needing massive numbers of examples to needing far fewer examples in some cases, is important for robotics. If a robot can learn useful behavior from a smaller and more diverse set of training cases, the path to practical deployment becomes less dependent on exhausting trial and error.
What Boston Dynamics Brings To The Deal
Boston Dynamics gives the collaboration a strong hardware platform. The company revealed the electric Atlas design in April after retiring the larger hydraulic version of the humanoid. Since then, public looks at the new robot have been limited, though a short video in August showed Atlas doing pushups.
Those demonstrations emphasized strength and movement. The larger question is what the robot can learn to do beyond controlled physical demos. Boston Dynamics has worked on software and AI for its own systems, but teaching robots to complete complex tasks with full autonomy is a different kind of problem.
Boston Dynamics CEO Robert Playter framed the partnership around that larger ambition. He said the companies want to accelerate general-purpose humanoid development and work on complex challenges that lead to useful robots for real-world problems.
The partnership is also notable because Boston Dynamics and TRI are tied to major automotive companies. Boston Dynamics is run by Hyundai, while TRI is connected with Toyota. The two automakers are direct competitors, yet the robotics work brings their research ecosystems into the same project.
The Race For Humanoid Robot Intelligence
Boston Dynamics is not alone in the humanoid robot space. Its competitors include Agility, Figure, and Tesla. According to the source, those companies have mainly chosen to build their AI teams internally rather than through a partnership like the Boston Dynamics-TRI deal.
Boston Dynamics also has a research spinout called The AI Institute, formerly The Boston Dynamics AI Institute. It is run by Boston Dynamics founder and former CEO Marc Raibert, but it remains independent from Boston Dynamics itself. The institute is also described as a significantly younger organization that is still building out its team.
TRI, meanwhile, has become less invested in the hardware side of robotics. That makes the match clearer: Boston Dynamics can provide the humanoid platform, while TRI can contribute research around learning and behavior models.
The Hard Part Is General Purpose
The long-term goal is a true general-purpose machine. In plain terms, that means a robot that can learn and do many of the things a person can do, and possibly more. That is far more difficult than building a robot that performs a narrow demonstration or follows a limited set of programmed routines.
Robot hardware has been moving closer to the level needed for more sophisticated tasks. Atlas shows the physical side of that progress, while Spot has benefited from SDK support that expands the range of things robots can do. But broader artificial general intelligence remains a much harder challenge.
The Boston Dynamics and TRI collaboration does not mean general-purpose humanoids are solved. It shows where the field is focusing its energy: combining capable robot bodies with learning systems that may reduce the amount of training needed for useful behavior.
For Atlas, the next stage is not only about movement. It is about whether AI-based robotic intelligence can help a humanoid robot move from impressive demonstrations toward practical, flexible work in real environments.