Vision-language-action (VLA) policies — models that translate visual and linguistic input into robot motion — frequently fail when a robot's executed movement deviates from the action it was commanded to perform. According to the paper, these execution errors stem from the robot's mechanics and operating conditions, including wear and changes in payload.
To address this, the authors introduce self-compensating VLA, a deployment-time adaptation method that lets a policy pre-compensate for a robot's execution errors as it generates commands. The approach requires no task rewards or labels; instead, it updates the policy online using the residual between the action commanded by the VLA and the motion the robot actually executes. The accompanying summary notes that small LoRA adapters are updated using proprioceptive feedback while the robot controller is left unchanged.
The team also presents RoboStress, a controlled simulation benchmark designed to stress-test VLA robustness across execution conditions that would be impractical to cover with physical robots alone. It combines established joint-level models of friction, backlash, compliance, and gravity-compensation error into seven deployment scenarios whose execution errors depend on the robot's state and motion history. The scenarios cover heavy payloads, thermal drift, and mechanical wear.
On RoboStress, self-compensating VLA reportedly achieves higher average task success than both the base policies and methods that build robustness in during training, outperforming domain randomization and RobustVLA across all seven scenarios with both backbones.
On two physical Piper arms with different usage histories — one new and one used for a year — the method raised average task success by more than 30 percentage points on each arm with both π₀ and π₀.₅, according to the summary. The gains extended to objects not seen in the task demonstrations, with π₀.₅ success on unseen objects rising from 16% to 64%.