Presentation
Talk: A Bottom-Up Approach to Intelligence
DescriptionThe rise of Foundation Models, including LLMs and VLMs, has brought us closer to creating AI agents that understand and act in the world. However, challenges remain, particularly in the "act" component, due to the scarcity of human motion data and a lack of action labels. This talk focuses on building intelligence from the ground up. Our main goal has been to achieve zero-shot sim-to-real transfer on hardware with minimal human intervention by integrating reinforcement learning and imitation learning into a streamlined motion imitation framework. We began by focusing on single-task policies and later expanded to multi-task policies capable of generalizing across behaviors. This talk will showcase unprecedented natural behaviors in dynamic tasks performed by the Boston Dynamics e-Atlas robot, marking a major advance in bridging the gap between human characters in graphics and physical humanoid robots in robotics.
Presenter
Event Type
Frontiers
TimeMonday, 11 August 20258:00am - 8:45am PDT
LocationWest Building, Rooms 301-305
Research & Education
Livestreamed
Recorded
Animation
Physical AI
Robotics
