Foundation Reward Models for General Robot Skill Acquisition

Explore the cutting-edge research on foundation reward models for general robot skill acquisition in this insightful talk by Jason Ma from Montreal Robotics. Delve into the challenges of training foundation models for robotics and discover innovative approaches to overcome the lack of internet-scale data in the field. Learn how to leverage in-the-wild human videos to supervise general-purpose value functions, enabling zero-shot specification of dense rewards for robot tasks. Examine the potential of value pre-training paradigms to encompass goal specification in various modalities, including language. Gain valuable insights into how pre-trained value models can serve as effective subgoal decomposers, facilitating long-horizon manipulation tasks without human task labels. Uncover the latest advancements in embodied intelligence and their implications for developing adaptable, general-purpose robots capable of learning and refining skills in unstructured environments.

Jason Ma: Foundation Reward Models for General Robot Skill Acquisition