Event
ARC: Human Corrections Guide Residual Correction of Robot Policies
1 reports · 1 sources
Anukriti Singh and co-authors propose Adaptive Residual Correction (ARC), a post-training module: on top of a frozen imitation policy, it uses reinforcement learning to learn residual corrections, and uses sparse human corrections to decide when to enable residual control and in which direction of the residual action space to explore. The authors report that on three precision manipulation tasks with real robots, ARC raises the average task success rate from 33.3% to 75.0%; human corrections define a 2-3 dimensional task-relevant subspace within the seven-joint residual space, preserving 96.0-97.7% of the squared correction magnitude. These results are as stated in the paper and await independent verification.
Report timeline
- ARC: Using Human Corrections to Guide Residual Correction of Robot Policies
The paper proposes Adaptive Residual Correction (ARC), which learns residual corrections via reinforcement learning on top of a frozen imitation policy, using sparse human corrections to decide when to activate residual control and in which direction of the residual action space to explore. On three precision manipulation tasks with a real robot, ARC raises average task success from 33.3% to 75.0%; the human corrections define 2-3 dimensional task-relevant subspaces of the seven-joint residual space, retaining 96.0-97.7% of the squared correction magnitude.
arXiv cs.RO ·