Teaching Robots New Behaviors

In this talk

1. 0:10

   Diffusion policy for dexterous skills

   Toyota Research Institute says its diffusion policy can teach robots skills with fewer demonstrations and less time. The team sees this as a step toward large behavior models that let robots perform a wider range of useful work.

2. 0:53

   Beyond pick and place

   The approach has enabled robots to use tools, pour liquids, and peel vegetables without new code or explicit skill programming. TRI says it has taught its robot fleet more than 60 different behaviors.

3. 1:30

   Teaching through teleoperation and touch

   A human demonstrates skills through teleoperation, after which the policy learns over several hours. Haptic feedback helps teachers guide the robots, and touch also improves the robots’ own learning: they struggled to flip a pancake without sensing contact with the unseen surface.

4. 2:20

   Building a curriculum for robots

   TRI aims to train robots on enough foundational skills to generalize to tasks they have not seen before. It planned to teach hundreds of behaviors by the end of 2023 and more than a thousand by the end of 2024, while using simulation and fleet learning to expand training.

5. 3:02

   A rapid pace of progress

   The researcher says the robots are showing dexterity beyond what they expected a year earlier. New behaviors are appearing frequently, sometimes from one day to the next.