Toyota's PUNYO 1.0 AI robot utilizes a large behavior model architecture and a generative AI diffusion policy to learn and perform complex tasks. The robot's behavior is generated using a method called Diffusion Policy, which is a generative AI technique that enables the robot to acquire complex, dexterous behaviors with unprecedented speed.

The Diffusion Policy method leverages a large behavior model (LBM) framework developed by Toyota Research Institute (TRI) in partnership with Professor Song's group at Columbia University. This framework allows the robot to adapt, learn new skills, and interact dynamically with its environment. The LBM is honed through haptic feedback from human teachers, along with goal-oriented linguistic cues. By utilizing the Diffusion Policy, the robot rapidly internalizes new behaviors, achieving operational consistency and efficiency.

The Diffusion Policy method offers several advantages, including applicability to multi-modal demonstrations, high-dimensional action spaces, and the ability to avoid inconsistent or erratic behavior. It allows human demonstrators to teach behaviors in a natural manner without worrying about confusing the robot. The method also enables planning in time, which helps the robot avoid inconsistent or erratic behavior.