The Toyota Research Institute (TRI) is making significant strides in the field of robot learning, showcasing advancements in research that can teach a robot a new skill literally overnight. Traditionally, machine learning required millions of training cases to teach robots, but the system developed by TRI shows that only dozens of training cases are needed to teach a robot a new skill. TRI has trained robots on 60 skills and counting using this method.
The system developed by TRI combines traditional robot learning techniques with diffusion models, similar to the processes used in generative AI models. This allows robots to learn skills that can function in diverse settings, overcoming the challenge of robots struggling to function in less-structured environments. The goal is to create robots that can adapt to changes in their environment, making them ideal for tasks like helping older people live independently.
TRI starts by teaching the system through teleoperation, where the robot is made to repeat a task over and over. This process takes a couple of hours, during which the system learns the task by replicating the actions of a human operator. The system uses all available data, including sight and force feedback, to produce a fuller picture of the task. Force feedback is particularly important for tasks that require precise manipulation, such as holding a tool correctly.
After the initial training, the system’s neural networks continue training overnight to fully learn the skill. TRI relies on diffusion policy, a new way of generating robot behavior, to represent a robot’s visuomotor policy as a conditional denoising diffusion process. This approach allows the robot to find meaning in randomized data, enhancing its ability to learn new skills.
These advancements in teaching robots new skills overnight have the potential to greatly enhance the capabilities and adaptability of robots in various real-world scenarios.
Sources:
– Toyota Research Institute (TRI) showcase at TechCrunch Disrupt’s Hardware Stage
– TRI CEO and Chief Scientist Gill Pratt
– TRI Senior Research Scientist Benjamin Burchfiel