NYU Researchers Find Consistency Key for Robot Dexterity

NYU Tandon and Robotics and AI Institute researchers show robot dexterity learning depends on synthetic training data consistency. Their study in IEEE Robotics and Automation Letters reveals rapidly exploring random trees (RRTs) generate inconsistent demonstrations that hinder imitation learning. The paper won the IEEE RA-L Best Paper Award. Robots learn better when motion-planning algorithms produce stable, repeatable virtual demonstrations. This finding shifts focus from human demonstrations to synthetic data quality in robot training.