RoboCat: A Self-Improving Agent for Robotic Manipulation - 2023 Fall Robotics Colloquium

Explore a cutting-edge robotics colloquium featuring Coline Devin from Google DeepMind as she presents "RoboCat: A Self-Improving Agent for Robotic Manipulation." Delve into the potential of leveraging diverse robotic experiences to master new skills and embodiments quickly. Learn about RoboCat, a multi-embodiment, multi-task generalist agent for robotic manipulation, designed as a visual goal-conditioned decision transformer. Discover how this innovative agent can generalize to new tasks and robots, both zero-shot and through rapid adaptation. Gain insights into the agent's capabilities through large-scale evaluations in simulation and on real robot embodiments. Understand how RoboCat's training data diversity contributes to cross-task transfer and improved efficiency in adapting to new tasks. This 51-minute talk, part of the Paul G. Allen School's 2023 Fall Robotics Colloquium, offers a glimpse into the future of robot learning and autonomous improvement loops in robotic manipulation.