Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing
Sohyun Lee, Yoonjae Baek, Jaesang Won, Jinnyeong Kim, Kang Hyunwoo, Seung-Hwan Baek, Ivan Laptev, Suha Kwak
POSTECH · MBZUAI
Published September 29, 2026 · on the feed October 7, 2026 · ▲ 40 · arXiv 2609.37334
The abstract
Vision-language-action (VLA) policies often fail when a robot's executed motion deviates from their commanded action. Such execution errors arise from the robot's mechanics and operating conditions, such as wear and payload changes. We propose self-compensating VLA, a deployment-time adaptation method that enables a VLA policy to pre-compensate for the robot's execution errors when generating commands. Without task rewards or labels, it updates the policy online using the residual between the action commanded by a VLA and the motion executed by the robot. To stress-test VLA robustness across execution conditions that are impractical to cover with physical robots alone, we introduce RoboStress, a controlled simulation benchmark. 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. On RoboStress, self-compensating VLA achieves higher average task success than both the base policies and methods that build in robustness during training. On two physical robot arms with different usage histories, it raises the average task success rate by more than 30 percentage points on each arm, and the gains extend to objects not seen in the task demonstrations.
The Brief
2:34A short podcast summary of the paper. Written and voiced by AI.
Transcript
Taming VLAs under Robot Execution Errors. Self-Compensation and Stress Testing. By Sohyun Lee and seven others, from the Pohang University of Science and Technology and MBZUAI. Published on the twenty-ninth of September. No code repository is linked to it.
A vision-language-action policy issues commands, and the robot is supposed to carry them out. Often it does not, quite. The motion the robot actually executes deviates from the action it was commanded to take, and these execution errors come from the robot's own mechanics and operating conditions: wear, for instance, or a change in payload. When that happens, VLA policies often fail. Testing for it is hard too, because covering the full range of execution conditions with physical robots alone is impractical.
Self-compensating VLA is a deployment-time adaptation method. It lets a VLA policy pre-compensate for the robot's execution errors at the moment it generates commands, so that what the robot ends up doing is closer to what was intended. It needs no task rewards and no labels. The only signal is the residual between the action the VLA commanded and the motion the robot executed, and the policy is updated online from that residual.
To stress-test robustness, the authors introduce RoboStress, a controlled simulation benchmark. It combines established joint-level models of four effects: friction, backlash, compliance, and error in gravity compensation. These are assembled into seven deployment scenarios, and in each the execution errors depend on the robot's state and on its motion history, rather than being fixed noise.
On RoboStress, self-compensating VLA achieves higher average task success than the base policies, and also higher than methods that build robustness in during training. On two physical robot arms with different usage histories, it raises the average task success rate by more than thirty percentage points on each arm. The gains also extend to objects that never appeared in the task demonstrations.
The method treats the mismatch between command and motion as information rather than as something to be trained away in advance. Each robot wears and carries loads differently, and learning from that particular robot's residuals while it works is what lets one policy adapt to arms with different histories.
The two halves of the paper support each other. Physical robots cannot cheaply be made to exhibit every kind of mechanical fault, so RoboStress supplies controlled, state-dependent errors in simulation. The physical arms then confirm that what helps there also helps on real hardware, and that the benefit carries over to objects the demonstrations never showed.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A policy can be right and the robot still wrong. Here is the gap.
You deploy a robot policy that turns camera images and instructions into motion commands. On the real arm, the motion it actually performs drifts from what was commanded.
The road ahead
- 1 · The problem ~7 min A robot arm does not move exactly as its policy commands, and each arm drifts in its own way, so a fixed policy keeps failing.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min The policy keeps learning on the robot, from the gap between what it commanded and what the arm's own sensors say happened.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Command, measure, compute the residual, build a corrected target, update small adapter weights, and let the next commands pre-compensate.
- 6 · The experiment ~6 min You’ll guess: On two physical robot arms with different usage histories, by how many percentage points did self-compensating VLA raise average task success?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.