Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing
This arXiv preprint introduces self-compensating VLA, a deployment-time adaptation method that updates a robot policy online using the residual between commanded and executed motion, plus RoboStress, a simulation benchmark for joint-level execution errors. On two physical Piper arms, it improves average task success by more than 30 percentage points per arm, and on RoboStress it beats base policies and training-time robustness methods. The work matters because mechanical wear, payload, and thermal drift cause real execution errors that standard VLA training does not address.
Can a vision-language-action policy adapt online, without task rewards or labels, to robot-specific execution errors that arise after commands are issued, and can those errors be systematically stress-tested?
When a VLA is deployed, the motion a robot executes often differs from the action the policy commanded because of friction, backlash, compliance, gravity-compensation mismatch, wear, payload, and temperature. Existing robustness work mostly injects synthetic noise into the policy's output commands and does not model state- and history-dependent joint-level physical errors, so policies degrade under real mechanical conditions.
Prior VLA robustness work focused on visual corruptions, distractors, and paraphrased instructions, or added synthetic Gaussian/adversarial perturbations to end-effector commands during training, such as domain randomization and RobustVLA. Simulated evaluations used direct action noise rather than physically grounded joint-level errors. These methods produce inconsistent gains and do not target per-robot mechanical variation.
Self-compensating VLA computes the residual between the commanded action chunk and the proprioceptively measured executed motion, constructs a pseudo-target by subtracting that residual, and updates only rank-4 LoRA adapters on the action expert with a flow-matching objective plus an anchor regularization. The base policy is frozen and the low-level controller is unchanged. RoboStress injects established joint-level models—Stribeck friction, gravity-compensation error, Tao-Kokotović backlash, and Spong compliance—into the control pipeline across seven deployment scenarios with fixed or linearly increasing severity.
On RoboStress averaged over seven non-clean scenarios, self-compensating VLA with π0.5 reaches 56.6% success vs 49.1% base, 51.0% domain randomization, and 49.8% RobustVLA; with π0 it reaches 44.2% vs 41.7% base, 39.8% DR, and 38.8% RobustVLA. On two physical Piper arms, with π0.5 Robot A improves from 43% base and 52% RobustVLA to 76% Ours, and Robot B from 43% base and 50% RobustVLA to 74% Ours. With π0, Robot A improves from 40% base and 47% RobustVLA to 72% Ours, and Robot B from 38% base and 49% RobustVLA to 71% Ours. On unseen object variants (25 episodes), π0.5 Ours achieves 64% vs 16% base and 36% RobustVLA. A disturbance observer plus adaptation achieves 80% vs 56% DOB alone on one task.
The authors state that the method relies on proprioception and propose using external motion sensors in future work. Evaluation is limited to two Piper arms, four real tasks, and 25 episodes per task; RoboStress is simulation-only with idealized joint-level models. No code or benchmark release is described. Long-term online adaptation stability, safety under larger errors, and generalization to other robot types are not assessed.
Robot arm manufacturers, warehouse automation providers, and VLA-based manipulation startups could use online residual adaptation to compensate for per-robot mechanical variation without retraining. The method is demonstrated on physical Piper arms with fine-tuned π0/π0.5 and requires proprioception plus per-deployment compute for LoRA updates. Integration into commercial stacks is plausible in 1–3 years if safety validation, cross-robot generalization, and reproducible implementations are established.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.