AI Insight
Researchers developed Adaptive Residual Correction (ARC), a method that improves robot manipulation policies by learning targeted corrections only where needed, rather than retraining entire systems. The approach uses sparse human corrections to identify when and where a robot should adjust its actions, then uses reinforcement learning to refine those specific movements. Testing on three real-world precision tasks improved average success rates from 33.3% to 75.0%, with human corrections defining low-dimensional subspaces (2-3 dimensions out of 7 joint movements) that captured 96-97.7% of correction magnitude.
Why it matters
This approach could make robot training more efficient by fixing localized failures without costly full retraining, potentially accelerating deployment of robots in manufacturing, surgery, and other precision tasks. The method reduces the burden on human trainers by requiring only sparse corrections rather than extensive demonstrations.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: Imitation policies that reliably execute most of a manipulation trajectory can still fail in localized regions requiring last-mile precision. Updating the full policy or exploring residual actions everywhere treats this local failure as a global learning problem. We introduce Adaptive Residual Correction (ARC), a lightweight post-training module that uses reinforcement learning (RL) to learn a residual correction on top of a frozen policy while using sparse human corrections to structure autonomous exploration. Rather than treating human corrections as imitation targets, ARC uses pre-intervention states to determine ‘when’ residual control should become active and correction geometry to determine ‘where’ in residual action space autonomous refinement should explore; task reward determines ‘what’ residual is ultimately executed. Across three precision-manipulation tasks on a real robot, ARC improves average task success from 33.3% to 75.0%. The human corrections define task-dependent 2-3 dimensional subspaces of the seven-joint residual space that retain 96.0-97.7% of squared correction magnitude. Together, these results show that sparse human corrections can localize and constrain autonomous post-training without requiring the observed correction itself to serve as the final action target.
Source: When, Where, and What to Correct: Human-Guided Adaptive Residual Correction for Robot Policies