Building Adaptable Generalist Robots

Explore the cutting-edge advancements in deep robot learning and the challenges of building adaptable generalist robots in this 48-minute talk by Mengdi Xu from Montreal Robotics. Delve into innovative approaches for improving robot generalization, including in-context learning from demonstrations, unsupervised continual reinforcement learning, and leveraging large foundation models for embodied agents. Discover how these techniques enhance data efficiency, parameter efficiency, and robustness, enabling robots to acquire new motor skills and solve complex physical puzzles with creative tool use. Gain insights into the future of robotics as Xu, a Ph.D. student at Carnegie Mellon University, shares her research on building learning-based robots capable of reliably interacting with the unstructured real world and adapting to unseen tasks beyond their training sets.