Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing
The paper’s core claim is that VLAs can adapt at deployment time by learning from the gap between commanded actions and the robot’s actual motion.
The authors propose a self-compensating VLA that updates online without task rewards or labels. They also introduce RoboStress, a simulation benchmark for testing execution errors from friction, backlash, compliance, and gravity-compensation problems. In RoboStress, the method outperforms base policies and training-time robustness methods. On two real robot arms, it improves average task success by more than 30 percentage points on each arm, including gains on unseen objects. HF Daily Papers' note
The authors propose a self-compensating VLA that updates online without task rewards or labels. They also introduce RoboStress, a simulation benchmark for testing execution errors from friction, backlash, compliance, and gravity-compensation problems. In RoboStress, the method outperforms base policies and training-time robustness methods. On two real robot arms, it improves average task success by more than 30 percentage points on each arm, including gains on unseen objects. HF Daily Papers' note
score 5