Do We Really Need All That Data? Learning and Control for Contact-rich Manipulation
Explore a thought-provoking seminar on data efficiency in robotic manipulation presented by Michael Posa, Assistant Professor of Mechanical Engineering & Applied Mechanics at the University of Pennsylvania. Delve into the challenges of adapting robots to new environments and tasks without extensive pre-training. Examine the clash between contact-driven aspects of manipulation and standard learning methods' inductive biases. Discover how contact-inspired implicit learning and convex optimization can reshape loss landscapes, leading to more accurate training and better generalization. Learn about the latest developments in deploying learned models through real-time multi-contact Model Predictive Control (MPC) for robotic manipulation. Gain insights into the potential for robots to gather information quickly and accomplish complex tasks in novel situations.