When, Where, and What to Correct: Human-Guided Adaptive Residual Correction for Robot Policies
Sparse human corrections helped raise average robot task success from 33.3% to 75.0% across three precision-manipulation tasks.
Adaptive Residual Correction adds a reinforcement-learning module on top of a frozen robot policy. Human interventions identify when correction should activate and constrain where the module explores; task rewards determine the final correction. The human corrections define task-dependent, two- or three-dimensional subspaces within the seven-joint residual space, retaining 96.0–97.7% of squared correction magnitude. The reported results come from tests on a real robot.
ArXiv · AI/CL/LG's note
Adaptive Residual Correction adds a reinforcement-learning module on top of a frozen robot policy. Human interventions identify when correction should activate and constrain where the module explores; task rewards determine the final correction. The human corrections define task-dependent, two- or three-dimensional subspaces within the seven-joint residual space, retaining 96.0–97.7% of squared correction magnitude. The reported results come from tests on a real robot.
ArXiv · AI/CL/LG's note
score 4