Robots That Know What to Ask: Recovering Misaligned Rewards through Targeted Explanations
Researchers have developed a new framework to help robots learn reward functions more accurately from human demonstrations. The system identifies underspecified features in demonstrations by analyzing the variation in behavior, indicating where the robot needs more guidance. It then prompts users for targeted corrective demonstrations, significantly improving reward recovery and reducing misalignment compared to random querying or passive data collection. AI
IMPACT Improves robot learning from human demonstrations by enabling targeted feedback, reducing misalignment.