Toyota Research Institute Unveils Large Behavior Models, LBMs

Toyota Research Institute (TRI) unveiled a new Generative AI teaching technique, Large Behavior Models (LBMs), enabling robots to adapt, learn new skills, and interact dynamically with its environment. Leveraging what's termed as "Diffusion Policy," the cutting-edge approach enables robots to acquire complex, dexterous behaviors with unprecedented speed. In other words, a robotics parallel to the transformative Large Language Models (LLMs) in conversational AI.

TRI's aim is to augment human abilities, and pave a way for robots to become more capable extensions of human endeavors.

Conventional techniques for teaching robots are cumbersome, limited to rigid environments, and often entail extensive coding or iterative trials. TRI's new approach has already empowered robots to master over 60 intricate skills, such as pouring liquids and manipulating soft materials. The entire teaching process bypassed the need for new code, relying solely on fresh data inputs. Their team aspires to scale that number to hundreds by year-end and reach 1,000 skills by 2024.

The TRI-developed robot behavior model is honed through haptic feedback from human teachers, in conjunction with goal-oriented linguistic cues. Utilizing its Diffusion Policy, the AI rapidly internalizes the new behavior, achieving operational consistency and efficiency in record time.

Developed in partnership with Professor Song's group at Columbia University, the "Diffusion Policy" is a Generative AI framework designed to expedite the robot's learning process - hence Large Behavior Models. TRI's robots themselves are custom-built with a focus on dual-arm dexterous manipulations and are equipped with advanced haptic and tactile feedback systems. "Drake," an open-source, model-based toolkit, serves as the backbone of TRI's robotics platform, enabling rapid development in both simulated and real-world settings.

Additionally, adhering to a "Safety First" philosophy, TRI has integrated stringent safety protocols to prevent the robots from colliding with themselves or their environment.