Building Blocks of Generalizable Autonomy

Explore the foundations of generalizable autonomy in robotics through this seminar presented by Animesh Garg, CIFAR Chair Assistant Professor at the University of Toronto. Delve into key aspects of machine learning for perception and control in robotics, focusing on developing reusable cognitive concepts and dexterous skills across various task instances. Examine three crucial areas: representational biases for embodied reasoning, causal inference in abstract sequential domains, and interactive policy learning under uncertainty. Discover how structured biases in modern reinforcement learning algorithms can be applied to robotics, covering state, actions, learning mechanisms, and network architectures. Investigate the discovery of latent causal structure in dynamics for planning, and learn how large-scale data generation combined with structure learning insights can enable sample-efficient algorithms for practical systems. Gain insights into applications in manipulation, surgical robotics, and legged locomotion from this comprehensive exploration of generalizable autonomy.