Taming VLAs

under Robot Execution Errors: Self-Compensation and Stress Testing

Scroll

arXiv 2026

Sohyun Lee*POSTECH

Yoonjae Baek*POSTECH

Jaesang WonPOSTECH

Jinnyeong KimPOSTECH

Hyunwoo KangPOSTECH

Seung-Hwan BaekPOSTECH

Ivan LaptevMBZUAI

Suha KwakPOSTECH

* Equal contribution

Paper arXiv Videos BibTeX

Robots don’t move exactly as commanded.
Let the VLA learn from that gap, while it runs.

Self-compensation at deployment.

Online LoRA updates from the command–execution residual. No rewards, no labels.

+30pp on real robots.

On a new and a 1-year-old Piper arm, for both π₀ and π_0.5.

Generalizes to unseen objects.

16% → 64% success on bowls and plates absent from fine-tuning.

RoboStress benchmark.

Four joint-level error models, seven deployment scenarios in simulation.

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.

Same robot, same VLA

A year-old Piper arm (Robot B). Left: the base policy. Right: the same policy with self-compensation.

“Pick up the gray bowl on the plastic cabinet and place it on the plate”π₀

Base

Ours

“Open the top drawer of the cabinet”π_0.5

Base

Ours

Learning from the gap

The executed motion misses the commanded target; after an online update, the compensated command lands on target.

Wear, payload, and heat make the executed motion drift from the command. That drift is visible through proprioception alone, so it can serve as a free training signal during deployment.

The pretrained VLA stays frozen. Only small LoRA adapters on the action expert are updated, and the robot’s low-level controller is left untouched.

01 MEASURE

Residual from proprioception

After each executed step, compare the achieved motion with the command. No reward, label, or error model is needed.

02 COMPENSATE

A corrected target

Subtracting the residual from the command gives the command that would have produced the intended motion.

03 ADAPT

Online LoRA update

Rank-4 LoRA is trained toward that target during deployment, so later chunks pre-compensate.

Two real arms, four tasks

A new AgileX Piper arm (Robot A) and one used for a year (Robot B). Both show real command–execution mismatch (mean normalized residual 32.9% and 35.0%).

Robot Backbone Task

Base

Ours

Average success over 4 tasks

25 episodes per task · top-right: gain of Ours over Base

Per-task results

Task	Robot A						Robot B
	π_0.5			π₀			π_0.5			π₀
	Base	RVLA	Ours	Base	RVLA	Ours	Base	RVLA	Ours	Base	RVLA	Ours
1	48	52	84	44	52	76	44	56	80	36	56	60
2	36	60	76	40	28	80	36	48	72	32	52	84
3	48	48	76	44	68	72	56	52	84	52	44	84
4	40	48	68	32	40	60	36	44	60	32	44	56
Avg.	43	52	76	40	47	72	43	50	74	38	49	71

Tasks 1–3: pick up the gray bowl (next to the plate / on the plastic cabinet / on the gift box) and place it on the plate. Task 4: open the cabinet’s top drawer. RVLA = RobustVLA.

Unseen objects

Bowls and plates never seen in the fine-tuning demonstrations. Robot A, π_0.5, 5 variants × 5 episodes.

Unseen wooden bowl: start and end frames

Unseen plum-colored bowl: start and end frames

Unseen light-blue bowl: start and end frames

RoboStress

A fleet of worn, hot, or heavily loaded robots is impractical to assemble. RoboStress simulates them with four joint-level error models, each injected where it physically arises.

Friction

Stribeck model: sticky at low speed.

torque

Gravity comp.

Pose-dependent load error.

torque

Backlash

Dead zone on direction reversal.

position

Compliance

Joints deflect under load.

position

EE command→ OSC→ torque + friction, gravity→ position + backlash, compliance→ physics

Seven deployment scenarios

Same LIBERO episode, clean (top) vs. under the scenario (bottom).

Clean

RoboStress

Scenario definitions

Scenario	Physical cause	Fric.	Grav.	Back.	Comp.	Severity over episode	Joints
Heavy Payload	grasped load		✓		✓	while grasping	all
Thermal Drift-Stribeck	temperature	✓				linearly increasing	all
Thermal Drift-Backlash	temperature			✓		linearly increasing	all
Aged Transmission	gear wear	✓		✓		fixed	all
Aged Joint-Uniform	gear + bearing wear	✓		✓	✓	fixed	all
Aged Joint-Shoulder	gear + bearing wear	✓		✓	✓	fixed	J1–J3
Aged Joint-Elbow	gear + bearing wear	✓		✓	✓	fixed	J4

Closer to real errors than Gaussian noise

End-effector deviation replaying a pick-and-place with a 1 kg bowl. Shared y-axis; shaded = before grasp; thin = repeats, thick = mean.

Real Piper arm holding a weighted bowl — Real

Real robot

reference

RoboStress

DTW to real 0.027

Gaussian action noise

DTW to real 0.388

Numbers on RoboStress

Success rate over the four LIBERO suites (40 tasks, 20 episodes each). Training-time robustness (DR, RobustVLA) gives inconsistent gains; self-compensation is best in all seven scenarios on both backbones.

Citation

@article{lee2026taming,
  author  = {Lee, Sohyun and Baek, Yoonjae and Won, Jaesang and Kim, Jinnyeong and
             Kang, Hyunwoo and Baek, Seung-Hwan and Laptev, Ivan and Kwak, Suha},
  title   = {Taming {VLAs} under Robot Execution Errors: Self-Compensation and Stress Testing},
  journal = {arXiv preprint arXiv:2609.37334},
  year    = {2026},
}