Towards Generalizable Autonomy - Duality of Discovery & Bias

Explore the frontiers of embodied intelligence and robotics in this Stanford seminar featuring Animesh Garg from Georgia Tech/NVIDIA. Delve into the challenges of generalization in interactive learning across task families, focusing on the development of efficient representation and inference mechanisms. Examine key questions in embodied AI, including representational biases, causal inference for decision-making, perceptual representations, and scalable learning systems. Discover innovative approaches such as C-Learning for balancing speed and reliability, latent causal structure discovery for improved sample efficiency, and task graphs for hierarchical manipulation tasks. Learn about practical applications through the Roboturk platform for scaling structured learning in robot manipulation, and explore algorithms for deployment with safety constraints. Gain insights into the future of embodied AI and its implications for both perception and decision-making in robotics.