Preference Learning from Minimal Human Feedback for Interactive Autonomy

This 28-minute talk from Open Data Science explores techniques for training robots with minimal human feedback, addressing the critical obstacle of limited robotics datasets that has hindered breakthroughs in robot learning and interactive autonomy. Discover how active learning methods can make reinforcement learning from human feedback (RLHF) more data-efficient, and learn about an alternative approach using language corrections to improve both data and time efficiency. The presentation includes accessible slides and examines why robotics hasn't experienced the same transformative advances as natural language processing and computer vision despite similar algorithmic approaches.