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

cs.ROSep 29, 2026

Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing

Authors: Sohyun Lee, Yoonjae Baek, Jaesang Won, Jinnyeong Kim, Kang Hyunwoo, Seung-Hwan Baek, Ivan Laptev, Suha Kwak

Organizations: POSTECH · MBZUAI

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.

Figures & tables

Figure 1: Overview of our study. (a) Self-compensating VLA updates the policy online using command-execution residuals and generates subsequent commands that pre-compensate for execution errors. (b) The RoboStress benchmark combines four established joint-level models into controlled deployment scenarios in simulation to stress-test VLA robustness.

Figure 2: Overview of self-compensating VLA. A frozen VLA with trainable LoRA adapters predicts an action chunk, which the robot executes. The residual $\eta^{\textrm{obs}}$ between the commanded and executed motion forms a pseudo-target $a_{t}^{\mathrm{target}}$ , which the compensation objective $\mathcal{L}_{\textrm{comp}}$ uses to update only the LoRA adapters online during deployment, leaving the base policy frozen.

Figure 3: Overview of the joint-level noise injection. The operational space controller ( Khatib, 1987 ) converts end-effector commands into joint torques. Torque-level noise (friction, gravity-compensation error) is added to joint torques, and position-level noise (backlash, compliance) to joint positions before physics integration. The next observation is rendered from the updated state.

Scenario	Physical cause	Fric.	Grav.	Back.	Comp.	Severity over episode	Joints
Heavy Payload	grasped load		$\checkmark$		$\checkmark$	While grasping	all
Thermal Drift-Stribeck	temperature change	$\checkmark$				Linearly increasing	all
Thermal Drift-Backlash	temperature change			$\checkmark$		Linearly increasing	all
Aged Transmission	gear wear	$\checkmark$		$\checkmark$		Fixed	all
Aged Joint-Uniform	gear + bearing wear	$\checkmark$		$\checkmark$	$\checkmark$	Fixed	all
Aged Joint-Shoulder	gear + bearing wear	$\checkmark$		$\checkmark$	$\checkmark$	Fixed	J1–J3

Table 1: Deployment scenarios in RoboStress. Checked noise components have increased severity, while unmarked components retain their nominal settings.

	$\pi_{0.5}$				$\pi_{0}$
Scenario	Base	DR	RobustVLA	Ours	Base	DR	RobustVLA	Ours
Clean	97.0	97.1	96.6	95.8	91.1	90.5	92.0	92.6
Heavy Payload	44.1	42.1	38.9	56.8	42.3	32.0	30.9	42.8
Thermal Drift-Stribeck	54.1	54.6	55.9	60.5	46.1	46.3	46.0	50.6
Thermal Drift-Backlash	83.4	84.5	83.1	84.8	75.5	77.3	77.3	78.8
Aged Transmission	33.2	39.7	40.2	41.8	26.8	22.9	23.8	28.9

Table 2: Success rates (%) on RoboStress deployment scenarios, averaged over LIBERO tasks. Averages exclude Clean. Best results within each backbone are shown in bold for each row.

	Robot A						Robot B
	$\pi_{0.5}$			$\pi_{0}$			$\pi_{0.5}$			$\pi_{0}$
Task	Base	RobustVLA	Ours	Base	RobustVLA	Ours	Base	RobustVLA	Ours	Base	RobustVLA	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

Table 6: Real-world success rates (%) on two Piper arms. Tasks 1–3 involve picking up the gray bowl from next to the plate, on the plastic cabinet, and on the gift box, respectively, and placing it on the plate. Task 4 involves opening the cabinet’s top drawer. Best results are shown in bold.

Figure 5: Rollouts of self-compensating VLA (Ours) and the base policy (Base) on Tasks 2 and 4, for Robot A and Robot B. Green and red borders mark success and failure.

Figure 6: Generalization to unseen objects, evaluated on Robot A. Left: Our rollouts with the gray bowl (ID) and object variants (OOD). Right: Success rates on all OOD variants.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

	$F_{c,j,0}$	$F_{s,j,0}$	$v_{s,j}$	$\sigma_{v,j,0}$	$B_{j,0}$	$M_{j}^{\text{eff}}$	$D_{j,0}$	$K_{j,0}$	$\beta_{j,0}$
Joint	(Nm)	(Nm)	(rad/s)	(Nm $\cdot$ s/rad)	(rad)	(kg $\cdot$ m 2 )	(Nm $\cdot$ s/rad)	(Nm/rad)
J1	2.175	3.480	0.15	0.348	0.00087	0.70	90.73	6000	0.05
J2	2.175	3.480	0.15	0.348	0.00087	0.70	90.73	6000	0.10
J3	2.175	3.480	0.15	0.348	0.00087	0.70	90.73	6000	0.06
J4	2.175	3.480	0.15	0.348	0.00087	0.70	90.73	6000	0.10
J5	0.480	0.720	0.10	0.048	0.00130	0.20	39.60	4000	0.08

Appendix

Table A1: Nominal noise parameters per joint

j

, defined in § 4.1 .

Component	Applied to	Full setting
Stribeck friction ( $w^{f}$ )	$F_{c,j}$ , $F_{s,j}$ , $\sigma_{v,j}$ ( $v_{s,j}$ fixed)	$\times 2$
Backlash ( $w^{b}$ )	$B_{j}$	$\times 15$
Compliance ( $w^{c}$ )	$K_{j}=K_{j,0}/w^{c}_{j}$ , $D_{j}=D_{j,0}/\sqrt{w^{c}_{j}}$ ( $M_{j}^{\mathrm{eff}}$ fixed)	$K\div 5$ , $D\div\sqrt{5}$
Gravity ( $w^{g}$ )	$\beta_{j}=w^{g}_{j}\beta_{j,0}$	$\times 4$

Appendix

Table A2: Full setting of each severity factor of Eq. ( 7 ).

w^{f}_{j}

w^{b}_{j}

, and

w^{c}_{j}

model wear, and

w^{g}_{j}

models payload mismatch.

Component	Setting	Injected values (J1–J4 / J5–J7)
Stribeck	friction $\times 2$	$F_{c}\,{=}\,4.350/0.960$ Nm, $F_{s}\,{=}\,6.960/1.440$ Nm, $\sigma_{v}\,{=}\,0.696/0.096$ Nm $\cdot$ s/rad, $v_{s}\,{=}\,0.15/0.10$ rad/s
Backlash	$B\times 15$	$B\,{=}\,0.01305/0.01950$ rad ( $0.748^{\circ}/1.117^{\circ}$ )
Compliance	$K\div 5$ , $D\div\sqrt{5}$	$K\,{=}\,1200/800$ Nm/rad, $D\,{=}\,40.58/17.71$ Nm $\cdot$ s/rad, $M^{\mathrm{eff}}\,{=}\,0.70/0.20$ kg $\cdot$ m 2 ( $\zeta=0.70$ )
Gravity	$\beta\times 4$	$\beta=[0.20,\,0.40,\,0.24,\,0.40,\,0.32,\,0.32,\,0.20]$

Appendix

Table A3: Settings of noise components. Values are reported as J1–J4 / J5–J7 unless stated otherwise.

Scenario	Components with increased severity	Injected values
Thermal Drift-Stribeck	Stribeck (ramp)	$F_{c}$ , $F_{s}$ , and $\sigma_{v}$ ramp $\times 1.5\!\to\!\times 2.0$ over the episode: $F_{c}\,{=}\,3.26\!\to\!4.35$ / $0.72\!\to\!0.96$ Nm, $F_{s}\,{=}\,5.22\!\to\!6.96$ / $1.08\!\to\!1.44$ Nm, $\sigma_{v}\,{=}\,0.522\!\to\!0.696$ / $0.072\!\to\!0.096$ Nm $\cdot$ s/rad.
Thermal Drift-Backlash	Backlash (ramp)	$B$ ramps $\times 5\!\to\!\times 15$ : $0.00435\!\to\!0.01305$ / $0.0065\!\to\!0.0195$ rad.
Heavy Payload	Compliance + Gravity (while grasped)	$K\,{=}\,1200/800$ Nm/rad, $D\,{=}\,40.58/17.71$ Nm $\cdot$ s/rad, $M^{\mathrm{eff}}\,{=}\,0.70/0.20$ kg $\cdot$ m 2 ; $\beta_{j}$ (J1–J7): 0.20, 0.40, 0.24, 0.40, 0.32, 0.32, 0.20.
Aged Transmission	Stribeck + Backlash	all joints: $F_{c}\,{=}\,4.35/0.96$ Nm; $F_{s}\,{=}\,6.96/1.44$ Nm; $\sigma_{v}\,{=}\,0.696/0.096$ Nm $\cdot$ s/rad; $B\,{=}\,0.01305/0.01950$ rad.
Aged Joint-Uniform	Stribeck + Backlash + Compliance	all joints: the above Stribeck and backlash parameters plus $K\,{=}\,1200/800$ Nm/rad, $D\,{=}\,40.58/17.71$ Nm $\cdot$ s/rad.
Aged Joint-Shoulder	Stribeck + Backlash + Compliance	J1–J3 at the full setting; J4–J7 nominal.

Appendix

Table A4: Deployment scenarios. Values are reported as J1–J4 / J5–J7 unless stated otherwise. “Full setting” applies the factors of Table A2 , and “nominal” denotes the values of Table A1 .

Scenario	Residual (%)
Simulation (RoboStress)
Clean	23.4
Heavy Payload	48.6
Thermal Drift-Stribeck	47.5
Thermal Drift-Backlash	46.1
Aged Transmission	38.7

Appendix

Table C5: Residual magnitude across scenarios, computed as

\|\eta^{\text{obs}}\|_{2}/\|a^{\text{policy}}\|_{2}

\pi_{0}

rollouts and expressed as a percentage. The simulation rows use the end-effector residual in normalized action units and the real-robot rows use the joint-space residual of § B.6 in radians.

Scenario	w/o anchor	w/ anchor
Heavy Payload	38.2	42.8
Thermal Drift-Stribeck	50.4	50.6
Thermal Drift-Backlash	77.9	78.8
Aged Transmission	25.2	28.9
Aged Joint-Uniform	22.8	24.0
Aged Joint-Shoulder	35.5	39.2

Appendix

Table D6: Effect of the anchor loss on

\pi_{0}

. Success rates (%) per deployment scenario, averaged over the four LIBERO suites.

Figure D1: Success rate (%) of the base policy and self-compensating VLA with

\pi_{0.5}

across five severity levels (L1 to L5) for each deployment scenario (rows) and LIBERO suite (columns) in RoboStress, averaged over the same three seeds as Table 2 .

	Wear factors			Heavy Payload		Thermal Drift ramp end
Level	Stribeck $w^{f}$	Backlash $w^{b}$	Compliance $K\div$	$K\div$	$\beta\times$	Stribeck	Backlash
L1 (mild)	1.5	5	2	2	2	1.5 $\to$ 1.5	5 $\to$ 5
L2	1.625	7.5	2.75	2.75	2.5	1.5 $\to$ 1.625	5 $\to$ 7.5
L3	1.75	10	3.5	3.5	3	1.5 $\to$ 1.75	5 $\to$ 10
L4	1.875	12.5	4.25	4.25	3.5	1.5 $\to$ 1.875	5 $\to$ 12.5
L5 (full)	2	15	5	5	4	1.5 $\to$ 2	5 $\to$ 15

Appendix

Table D7: Severity levels. Each column moves linearly from L1 to L5, and L5 is the setting of Table 2 . Compliance divides the stiffness by the listed factor and the damping by its square root (

K_{j}=K_{j,0}/w^{c}_{j}

D_{j}=D_{j,0}/\sqrt{w^{c}_{j}}

), which keeps

\zeta=0.70

at every level. Wear factors apply to the worn joints of each scenario, that is, all joints for Aged Transmission and Aged Joint-Uniform, J1–J3 for Shoulder, and J4 for Elbow. Thermal Drift entries give the ramp as onset

\to

end multiplier.

Explore similar work

Self-Adaptive VLA for Robust Robot Deployment
Sep 24, 2026Hongxin Zhang, Chunru Lin, Tsun-Hsuan Wang +2Robot Systems
University of Massachusetts Amherst · Genesis AI
FATE-VLA:Failue-aware test generation for vision-language-action models
Jun 1, 2026Arusa Kanwal, Pablo Valle, Shaukat Ali +1Vision-Language-Action Framework Robot Policies
Mondragon University Mondragon, Spain · Simula Research Laboratory Oslo, Norway
VLAMotor: Test-Guided Enhancement of Vision-Language-Action Models via Agent-BasedData Synthesis
May 16, 2026Zeqin Liao, Peifan Ren, Zixu Gao +6Diffusion-Based Vision-Language-Actions Ava-Vlm
School of computing and data science, Nanyang Technological University · School of Software Engineering, Sun Yat-sen University, China and GuangDong Engineering Technology Research Center of Blockchain, China · Northwest A&F University

Sep 24, 2026, cs.ROJ/K move · Enter open · S save

Self-Adaptive VLA for Robust Robot Deployment

Hongxin Zhang, Chunru Lin, Tsun-Hsuan Wang, Zhenjia Xu+1

University of Massachusetts Amherst · Genesis AI

Robot Systems

Open paper AskPDF arXiv

While Vision-Language-Action (VLA) models demonstrate impressive capabilities in robotic manipulation, their memoryless nature renders them brittle to test-time environment shifts, particularly hardware shifts caused by wear or imperfect calibration. Enabling these models to self-adapt during deployment without requiring continuous on-site recalibration remains a critical bottleneck for real-world scalability. In this work, we introduce Self-Adaptive VLA, a novel post-training recipe that enables the policy to iteratively adapt to deployment-time hardware shifts leveraging its own rollouts as context. To do so, we first collect policy rollouts under deliberately injected hardware shifts. We then transform the base policy's training data into shift-conditioned expert demonstrations by pre-compensating the expert actions for these known shifts. Next, we introduce a lightweight, plug-in context encoder that compresses the context, including visual observation, proprioception, and actions in the shifted environment, into a latent context token. This token modulates the policy through adaptive layer normalization (AdaLN). Furthermore, we find that context tokens can be ensembled, allowing the policy to iteratively self-correct and mitigate failures step by step. Extensive experiments across four precision-critical bi-manual and dexterous manipulation tasks show that Self-Adaptive VLA recovers over 80% of the base policy's performance under hardware shifts, such as actuation bias and joint encoder offsets. Moreover, Self-Adaptive VLA enables more robust deployment to new workstations compared to the base policy. Our approach provides a pathway for robust large-scale real-world robot deployments and easier maintenance. See videos at https://icefoxzhx.github.io/self-adaptive-vla.

Abstract

Explore similar work

Self-Adaptive VLA for Robust Robot Deployment

FATE-VLA:Failue-aware test generation for vision-language-action models

VLAMotor: Test-Guided Enhancement of Vision-Language-Action Models via Agent-BasedData Synthesis