.webp)
TRI's new Diffusion Policy technique accelerates robot learning, aiming to reach 1,000 skills and enhance human-robot collaboration.
Toyota Research Institute (TRI) has revealed an innovative approach to teaching robots using Diffusion Policy, a method that builds large behavior models (LBMs) to rapidly impart new skills. Announced in September 2023, this advancement in generative AI represents a significant leap in robotics technology.
Unlike old-style AI methods that require extensive coding and iterative testing, TRI’s technique allows robots to learn from haptic demonstrations provided by human instructors. This approach extremely reduces training time, enabling robots to acquire new abilities in days instead of weeks. TRI's system has already taught robots over 60 tasks, including pouring liquids, using tools, and manipulating soft objects.
The new method utilizes the Drake model-based design, which offers a realistic simulation platform and an open-source optimization framework. By the end of 2024, TRI hopes to increase the robots' skill set to 1,000 activities, which will improve their capacity to carry out difficult, unstructured work in typical settings.
The head scientist at Toyota Motor Corp. and CEO of TRI, Gill Pratt, highlighted that the goal of this research is to enhance human talents rather than to replace them. TRI is laying the groundwork for more adaptive and helpful robotic systems by facilitating more effective interactions between robots and their environment.
.webp)
