Generative AI garnered a lot of attention last year, but developers have only just started applying it to robotics. In September 2023, Toyota Research Institute, or TRI, announced what it said was a “breakthrough” approach based on Diffusion Policy to quickly teach robots new skills.

The artificial intelligence advancement was a step toward building large behavior models (LBMs), which promise to do for robots what large language models (LLMs) have done for conversational AI, said TRI. “Our research in robotics is aimed at amplifying people rather than replacing them,” said Gill Pratt, CEO of TRI and chief scientist at parent company Toyota Motor Corp. “This new teaching technique is both very efficient and produces very high-performing behaviors, enabling robots to much more effectively amplify people in many ways.”

Previous AI approaches had limitations such as narrowly defined tasks and the need for numerous time-consuming iterative cycles, according to TRI, which has offices in Los Altos, Calif., and Cambridge, Mass.

In TRI’s approach, a teacher instructed a custom two-armed robot with haptic demonstrations, and the system then uses Diffusion Policy to learn the skill. It can do so with dozens of demonstrations in days rather than the weeks spent coding and testing, the company added.

The Drake model-based design provided TRI with a toolbox and realistic simulation platform. The optimization framework began as an internal model and has since been released as open-source.

TRI said it has taught robots to do more than 60 tasks such as pouring liquids, using tools, and manipulating soft objects by supplying them with new images and tactile sensing data rather than by writing code. TRI said it hopes to reach 1,000 skills by the end of 2024.

By enabling robots to interact with the world in more varied ways, Toyota Research Institute said it is making them better able to support people in everyday, unstructured environments.

Explore the RBR50 Robotics Innovation Awards 2024.