1. News
  2. Teaching robots new tricks, faster

Teaching robots new tricks, faster

Toyota Research Institute found a way to teach robots new, dexterous skills quickly and confidently
No, robots are not taking over the world. But soon they may be smart enough to take over simple mundane tasks such as, for example, cooking. And the ‘soon’ might be even sooner with the new breakthrough in the learning technology achieved by roboticists at the Toyota Research Institute. 

Game-changing new technology

A group of scientists at the TRI Robotics labs are working on improving robots so they can amplify people – in the spirit of Mobility for All. And their latest advancement is a game-changing tech that's making robots way smarter and more helpful. This advancement is based on an AI based Diffusion Policy, and in simple terms it allows robots to fast-track the way they learn new skills. The technology is a significant step towards building ‘Large Behavior Models (LBMs)’ for robots, in the same way that the ‘Large Language Models’ (LLMs) recently revolutionised conversational AI. 

Expanding learning capability

Unlike previous methods, which were slow and limited to specific tasks, this approach has already enabled our colleagues at TRI to teach robots more than 60 dexterous skills without writing a single line of new code. TRI's goal is to expand this capability further, with plans to teach hundreds of new skills by the end of the year and 1,000 by the end of 2024.
“This new teaching technique is both very efficient and produces very high performing behaviours, enabling robots to much more effectively amplify people in many ways.”
Said Gill Pratt, CEO of TRI and Chief Scientist for Toyota

New skill in less than a day

This is how it works in real life. A teacher instructs a robot to perform certain actions using a teleoperation system by essentially demonstrating a small set of skills. This marks the beginning of the process. Next, the AI-based Diffusion policy absorbs information and refines the robot's capabilities in the background over several hours. Typically, the teaching of the robot is done in the afternoon, and the learning process happens overnight. When the team returns in the morning, they find the robot able to execute new behaviours it learned during the night.

Learning through touch

A sense of touch is a vital aspect of this learning process. Just like humans learn better through touch and interaction, robots benefit greatly from it too. Thanks to a haptic device, which simulates the sense of touch for the teacher, robots can now learn and improve their skill by interacting with its environment. Otherwise, it would struggle to perform tasks effectively. When it can interact with its surroundings through touch, it becomes successful in carrying out various actions, like flipping a pancake, demonstrating the power of incorporating touch in the learning process.
“What is so exciting about this new approach is the rate and reliability with which we can add new skills. Because these skills work directly from camera images and tactile sensing, using only learned representations, they are able to perform well even on tasks that involve deformable objects, cloth, and liquids — all of which have traditionally been extremely difficult for robots.”
Said Russ Tedrake,Vice President of Robotics Research at TRI