LOS ALTOS, Calif. and CAMBRIDGE, Mass. (Sept. 19, 2023) – The Toyota Research Institute (TRI) has announced a significant advancement in robotics through a new generative AI technique called Diffusion Policy. This innovative approach dramatically accelerates the process of teaching robots complex and nuanced skills, marking a major step towards the development of “Large Behavior Models” (LBMs) for robots – the robotic equivalent of the Large Language Models (LLMs) that have transformed conversational AI.
“Our fundamental belief at Toyota Research Institute is that robotics should amplify human capabilities, not replace them,” stated Dr. Gill Pratt, CEO of TRI and Chief Scientist for Toyota Motor Corporation. “This groundbreaking teaching method is not only remarkably efficient but also results in exceptionally high-performance behaviors, empowering robots to become significantly more effective partners for people in a multitude of applications.”
Traditionally, training robots to perform new tasks has been a laborious and often unpredictable process. Existing state-of-the-art methods demanded extensive hours from roboticists to write intricate code or engage in numerous trial-and-error iterations to program even relatively simple behaviors. These conventional techniques were often constrained to narrowly defined tasks within meticulously controlled environments, limiting the real-world applicability of robots.
TRI’s new Diffusion Policy approach revolutionizes this paradigm. Researchers at the Toyota Research Institute have already successfully trained robots in over 60 intricate, dexterous skills, encompassing tasks such as pouring liquids with precision, effectively utilizing various tools, and skillfully manipulating deformable objects. Remarkably, these achievements were accomplished without writing any new lines of code; the sole modification was providing the robot with fresh datasets. Building upon this remarkable initial success, TRI has set ambitious goals to expand the robot’s repertoire to hundreds of skills by the close of the current year and an impressive 1,000 skills by the end of 2024.
This announcement underscores a pivotal shift in the capabilities of robots. No longer confined to basic “pick and place” actions, Toyota Research Institute robots are now being taught to operate effectively in diverse and dynamic scenarios, executing a broad spectrum of behaviors. This enhanced dexterity and adaptability will pave the way for robots to provide meaningful support to individuals in everyday situations and navigate the complexities of unpredictable, constantly evolving environments.
Dr. Russ Tedrake, Vice President of Robotics Research at TRI and also the Toyota Professor at MIT, expressed his astonishment at the rapid progress. “The dexterity and versatility I am witnessing in these robots are truly astounding. Even just a year ago, I would not have anticipated such a significant leap forward in such a short timeframe,” he commented. Dr. Tedrake further elaborated on the significance of this breakthrough, stating, “The truly exciting aspect of this new methodology is the speed and reliability with which we can integrate new skills. Because these skills are learned directly from visual input from cameras and tactile feedback, relying solely on learned representations, they exhibit exceptional performance even when dealing with tasks involving deformable objects, fabrics, and liquids – areas that have historically posed immense challenges for robots.”
Deep Dive into the Technology: Diffusion Policy and Robot Learning at TRI
The core of TRI’s innovative approach lies in its robot behavior model, which learns through haptic demonstrations provided by a human teacher, coupled with a language description outlining the desired goal. The robot then employs an AI-driven Diffusion Policy to master the demonstrated skill. This streamlined learning process enables the autonomous deployment of a new behavior after just a few dozen demonstrations. This method not only yields consistent, repeatable, and high-performing results but also achieves this with remarkable speed and efficiency, pushing the boundaries of robot learning.
Key Technological Achievements Fueling TRI’s Breakthrough:
-
The Power of Diffusion Policy: Developed in collaboration with Professor Song’s group at Columbia University, Diffusion Policy represents a cutting-edge generative AI technique for behavior learning. This method facilitates rapid and straightforward teaching of new behaviors through demonstration, significantly reducing the complexities of traditional robot programming.
-
Custom-Built Robot Platform: Toyota Research Institute has engineered a specialized robot platform explicitly designed for intricate dual-arm manipulation tasks. A key feature of this platform is its enhanced haptic feedback and tactile sensing capabilities, allowing for more nuanced and responsive interactions with the environment.
Alt text: Toyota Research Institute robot utilizing dual arms and tactile sensors to perform a complex manipulation task, showcasing dexterity achieved through Diffusion Policy learning.
-
Scalable Skill Pipeline: The effectiveness of TRI’s approach is evident in its rapid skill acquisition pipeline. Having already equipped robots with 60 dexterous skills, Toyota Research Institute is on track to reach its ambitious targets of hundreds of skills by year-end and a thousand by the close of 2024, demonstrating the scalability of their method.
-
Drake: A Foundation for Innovation: A critical component of TRI’s success is Drake, an open-source model-based design toolbox for robotics. Drake provides a state-of-the-art environment for both simulation and real-world development, dramatically increasing the speed and scale of research and development. TRI’s internal robot architecture is built upon Drake’s optimization and systems frameworks, and its open-source nature fosters collaboration and accelerates progress across the entire robotics community.
Alt text: Visualization of Drake simulation environment, highlighting its role in enabling rapid development and testing of robot behaviors at Toyota Research Institute.
- Safety as a Core Principle: Safety is paramount in all robotics endeavors at Toyota Research Institute. The system incorporates robust safeguards, powered by Drake and TRI’s custom robot control stack, to ensure robots operate safely, preventing collisions with themselves or their surroundings. This commitment to safety is integral to TRI’s responsible approach to robotics development.
Further technical details about Diffusion Policy are available in the research paper Diffusion Policy, published at the 2023 Robotics Science and Systems conference. Additional insights can be found on TRI’s Medium blog.
For those seeking deeper engagement, Toyota Research Institute hosted a LinkedIn Live Q&A session on October 4th. Recordings and further information may be available on TRI’s LinkedIn page.
This breakthrough by Toyota Research Institute not only signifies a major leap forward in robot learning but also holds immense potential for the future of human-robot collaboration, paving the way for robots to become truly helpful and versatile partners in our daily lives.
