The system relies on a diffusion policy to generate robot behavior by representing a robot's visuomotor policy as a conditional denoising diffusion process. The goal is to create Large Behavior Models to help robots learn and scale. The team hopes these findings will lead to more capable robots that can interact with novel objects in new settings and create actions based on trained behaviors.
Key takeaways:
- The Toyota Research Institute (TRI) is developing a system that can teach a robot a new skill overnight, using traditional robot learning techniques and diffusion models.
- TRI's method allows robots to function in diverse settings and adapt to changes, which is crucial for robots to operate in real-world environments.
- Roboticists at TRI teach the systems through teleoperation, a process that can take a couple of hours, wherein the system is made to repeat the same task over and over.
- TRI is also working on creating Large Behavior Models to help robots learn and is aiming to train thousands of skills by the end of next year.