Interactive Learning

Continuously improve robot policies through online human correction signals — where a human supervisor provides corrective feedback that the policy uses to improve without full re-training. Interactive learning combines the safety of imitation learning with the improvement capability of RL. A human supervisor watches the policy execute and provides corrective interventions or demonstrations when the policy makes mistakes (DAgger). These corrections are aggregated into the training dataset, and the policy is updated to avoid the corrected mistakes. This cycle continues until the policy achieves desired performance. Modern variants use only corrective feedback (not full demonstrations) and can handle high-frequency intervention.