Toyota, in collaboration with MIT and Columbia Engineering, has made a significant breakthrough in AI learning for robots, showcasing a method that dramatically accelerates the acquisition of new skills. The approach, analogous to a “ChatGPT moment for robotics,” aims to usher in the general-purpose robotics era, with humanoid robots autonomously navigating workspaces and taking over various tasks from humans.
While early use cases primarily involve robots handling physical tasks, such as lifting and moving objects, Toyota envisions a future where robots can quickly adapt and learn a wide array of tasks based on human instructions or demonstrations. The key to this breakthrough is Toyota’s new learning approach called Diffusion Policy, developed in partnership with MIT and Columbia Engineering.
Diffusion Policy enables robotic AIs to observe and learn from human actions in the real world, programming themselves to perform tasks flexibly. Unlike some approaches that use virtual reality telepresence, Toyota’s method focuses on haptics. Human operators receive haptic feedback from the robot’s soft, flexible grippers through hand controls, allowing them to feel the robot’s interactions with objects.
The learning process begins with a human demonstrating a set of skills through teleoperation. Toyota’s AI-based Diffusion Policy then learns in the background over a matter of hours, creating its internal model of task success and failure. The robot then runs thousands of physics-based simulations based on its internal models to refine its techniques for the task.
The team has successfully trained robots in over 60 small tasks, including kitchen-based activities like spreading a slice of bread evenly with a knife or flipping a pancake with a spatula. Toyota aims to have hundreds of tasks mastered by the end of the year, with a target of over 1,000 tasks by the end of 2024. The ultimate goal is to create the first Large Behavior Model (LBM), a comprehensive AI-generated model of how robots can interact with the physical world to achieve various outcomes.
While Toyota’s breakthrough showcases the rapid evolution of robotics, experts emphasize that it’s not a complete solution to the climate crisis. The technology represents a crucial step in enabling robots to quickly learn new tasks, potentially revolutionizing industries and applications. As the team constructs the LBM, it sets the stage for future robot owners and operators to rapidly teach their bots new skills, upgrading entire fleets of robots with unprecedented dexterity and adaptability.
By Impact Lab