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

• Execution errors from mechanics and operating conditions remain underexplored, while prior robustness studies mainly examine visual or instruction changes and synthetic output perturbations.Physical fleet-scale testing is impractical and varying conditions can risk hardware damage; synthetic perturbations do not explicitly model state, motion-history, or joint-specific effects.Source: §1 · pp. 1–2
• Self-compensating VLA uses command-execution residuals to adapt the policy online without task rewards, labels, or knowledge of error sources, then generates commands that pre-compensate for observed errors.The robot’s low-level controller remains unchanged.Source: §1 · pp. 1–2
• RoboStress combines joint-level models of friction, backlash, compliance, and gravity-compensation error into seven simulation scenarios with errors dependent on robot state and recent motion history.The models are applied at corresponding stages of the control pipeline.Source: §1 · p. 2, Figure 1 · p. 2
• On RoboStress, self-compensating VLA outperforms two base policies and training-time robustness methods across all seven scenarios; on both physical arms, it raises average task success by more than 30 percentage points.Gains also extend to objects not seen in task demonstrations.Source: §1 · p. 2

• Prior VLA robustness work examines input changes, including viewpoint, lighting, and instructions, or synthetic perturbations of end-effector commands; this work instead targets execution errors from robot mechanics and operating conditions.Related studies include Fei et al. (2026), Liu et al. (2025), Zhang et al. (2025a, 2025b), and Guo et al. (2026).Source: §2 · pp. 2–3
• RoboStress repurposes established joint-level friction, backlash, compliance, and gravity-compensation models, originally developed for identification and control, to generate execution errors for VLA evaluation.The passage cites Canudas de Wit et al. (1995), Tao & Kokotovic (1993), Spong (1987), and Ma & Hollerbach (1996).Source: §2 · p. 3
• Earlier deployment-time adaptation methods use auxiliary self-supervision or infer conditions from recent states and actions, while EVOLVE-VLA (Bai et al., 2025) and TT-VLA use task-progress feedback.Self-compensating VLA instead adapts from command-execution residuals measured through proprioception, without task rewards.Source: §2 · p. 3
• Model-based, disturbance-observer, adaptive, and learning-based controllers address robot control through dynamics, feedback, parameter estimation, or learned compensation; self-compensating VLA adapts policy commands sent to the existing controller.It does so without modifying that controller or requiring an explicit dynamics model.Source: §2 · p. 3

3 SELF-COMPENSATING VLA

• The method obtains a compensation signal from commanded-versus-executed motion residuals, then updates the policy online so subsequent commands pre-compensate for execution errors.The method adapts a pretrained VLA to an individual robot during deployment.Source: §3 · p. 3

3.1 COMPENSATION SIGNAL

• At each executed step, the residual compares the achieved end-effector displacement with the commanded delta pose, using proprioceptive measurements scaled to the command units.The residual provides a self-supervisory signal for policy adaptation.Source: §3.1 · p. 4

3.2 DEPLOYMENT-TIME POLICY ADAPTATION

• Because the residual is observed only after execution, it forms a compensated pseudo-target for updating the policy on subsequent chunks rather than correcting the current chunk.The target is constructed by subtracting the observed residual from the executed command.Source: §3.2 · p. 4
• Online adaptation updates only LoRA parameters while keeping pretrained weights frozen, using a flow-matching objective on target action chunks and an anchor penalty to limit parameter drift.Target chunks from executed chunks are accumulated in an online buffer for deployment-time updates.Source: §3.2 · p. 4

4 THE ROBOSTRESS BENCHMARK

• RoboStress combines joint-level models of friction, gravity-compensation error, backlash, and compliance into a controlled benchmark of execution errors.The benchmark groups these effects into seven deployment scenarios.Source: §4 · p. 4, §4.1 · p. 4

4.1 JOINT-LEVEL NOISE MODELS

• Friction follows a velocity-dependent Coulomb-Stribeck law with a viscous term, while gravity-compensation error is a configuration-dependent fraction of controller-estimated gravitational torque.Friction is elevated near zero speed; the gravity-error fraction depends on joint angles, not motion history.Source: §4.1 · pp. 4–5
• Backlash uses a deadband model in which motor motion can leave the link still during direction reversal, with link deviation depending on its previous angle.The backlash term represents angular deviation between motor and link and is history-dependent.Source: §4.1 · p. 5
• Compliance models load-induced link deflection with spring-mass-damper dynamics that evolve over time and can produce transient overshoot and oscillation.The deflection reflects load history rather than only the current load.Source: §4.1 · p. 5

4.2 JOINT-LEVEL NOISE INJECTION

• Friction and gravity-compensation error perturb controller torques, while backlash and compliance perturb joint positions before simulator integration.Forward dynamics propagate these effects across coupled joints, making end-effector errors depend on arm pose.Source: §4.2 · p. 5

4.3 DEPLOYMENT SCENARIOS

• The seven deployment scenarios vary component severity by joint and may keep it fixed or ramp it linearly within an episode.They combine noise components to represent conditions such as wear, warming, and heavy payloads.Source: §4.3 · pp. 5–6, Table 1 · p. 6
• Heavy Payload increases gravity-compensation error and joint deflection, whereas Thermal Drift variants ramp friction or backlash severity during an episode.The payload scenario keeps friction and backlash nominal; the drift variants test adaptation to changing errors.Source: §4.3 · p. 6
• Aged Transmission increases friction and backlash across joints, while Aged Joint variants also increase compliance and target all joints, shoulder joints, or the elbow.The shoulder variant affects J1–J3; the elbow variant affects J4.Source: §4.3 · pp. 6–7

5.1 EXPERIMENTAL SETTING

• Evaluation covers four LIBERO task suites with 10 tasks each, 20 episodes per task per RoboStress setting, and 25 episodes per task on two Piper arms.The experiments use π0 and π0.5 as base policies; the real-world policies are fine-tuned on teleoperated demonstrations collected on each arm.Source: §5.1 · p. 7

5.2 EXPERIMENTS ON ROBOSTRESS

• Average success reaches 56.6% from 49.1% with π0.5 and 44.2% from 41.7% with π0; DR and RobustVLA give inconsistent gains and fall below π0 on average.Self-compensating VLA outperforms the base policy, DR, and RobustVLA in all seven deployment scenarios with both backbones.Source: §5.2 · p. 7
• The adjusted Heavy Payload model captures the post-grasp rise in real-arm deviation and has lower dynamic time warping distance to the real mean profile than Gaussian action noise.The model parameters were adjusted to match end-effector deviations from an unloaded replay; the simulation comparison used ten initial layouts.Source: §5.2 · p. 7
• Average success with π0 rises from 50.0% to 51.5% across noise components, with the smallest gain under compliance, where the base policy performs best.Table 3 evaluates individual noise components.Source: §5.2 · p. 8, Table 3 · p. 8
• With π0, command-target online updates reach 40.8%, while fixed chunk-level residual correction reaches 23.6%; policy adaptation provides observation-conditioned compensation.The command-target variant updates online but uses policy commands as targets; direct residual correction keeps the policy fixed.Source: §5.2 · p. 8, Table 4 · p. 8

5.3 REAL-WORLD EXPERIMENTS

• On both Piper arms, self-compensating VLA outperforms base policies and RobustVLA on every task with both π0 and π0.5.The arms differ in usage history; even the new arm has a 32.9% mean normalized residual, compared with 35.0% on the year-used arm. Representative rollouts show successes where the base policy fails on Tasks 2 and 4.Source: §5.3 · p. 8, Table 6 · p. 8, Figure 5 · p. 9
• On Robot A Task 1 with π0.5, DOB raises success from 48% to 56%, and adding adaptation raises it to 80%, versus 84% for adaptation alone.The results indicate that policy adaptation remains effective alongside external command correction and improves on DOB alone.Source: §5.3 · p. 8, Table 5 · p. 8
• On Robot A, self-compensating VLA reaches 64% success on five object variants absent from fine-tuning, versus 16% for the base policy and 36% for RobustVLA.The evaluation updates the instruction for each variant and uses five episodes per variant; Figure 6 shows the rollout and OOD success-rate comparisons.Source: §5.3 · p. 9, Figure 6 · p. 9

6 CONCLUSION

• The work combines online residual-based policy adaptation without task rewards or labels with RoboStress, which integrates four joint-level error models into seven deployment scenarios.Experiments report improved task success over base policies and training-time robustness methods, with real-world gains extending to objects absent from fine-tuning demonstrations.Source: §6 · p. 9
• Future work includes incorporating external-sensor motion estimates to reduce reliance on proprioception.The paper states that code for self-compensating VLA and RoboStress, including scenario configurations and evaluation scripts, will be released.Source: §6 · p. 9, REPRODUCIBILITY STATEMENT · p. 10
• Reported task-success improvements do not constitute safety guarantees; human-environment deployment requires additional safety evaluation and appropriate safeguards.The experiments evaluate manipulation on two robot arms.Source: ETHICS STATEMENT · p. 10

A.1 NOISE PARAMETERS

• RoboStress’s nominal per-joint parameters cover friction, backlash, dynamic compliance, and gravity-compensation residuals; the residual fraction ranges from 5% to 10%.The listed largest residual fractions are assigned to shoulder pitch (J2) and elbow (J4).Source: Table A1 · p. 14, §A.1 · p. 14

A.2 SEVERITY FACTORS

• Severity factors scale noise parameters; compliance damping is also scaled to preserve a damping ratio of 0.70 as stiffness drops.The payload factor scales the gravity-compensation residual fraction; the listed factors are wf = 2, wb = 15, wc = 5, and wg = 4.Source: §A.2 · p. 14, Table A2 · p. 14

A.3 NOISE INJECTION DETAILS

• Noise is recomputed at each 2 ms physics sub-step, with torque components added to controller torque and position components applied as joint-state offsets.The two injection layers couple through compliance; observed residuals reflect effects through arm dynamics and kinematics, not the injected noise itself.Source: §A.3 · pp. 14–15

A.4 SETTINGS OF EACH NOISE COMPONENT

• Each Table A3 setting activates one noise component across all joints and disables the other three.The settings specify component values separately for joints J1–J4 and J5–J7; the gravity setting uses β × 4.Source: §A.4 · p. 15, Table A3 · p. 15

A.5 DEPLOYMENT SCENARIOS

• The seven deployment scenarios combine noise components, including linearly ramped Thermal Drift variants, payload-gated Heavy Payload, and aged-joint cases targeting different joints.In the three Aged Joint cases, wear is applied uniformly, to J1–J3, or to J4; unlisted components remain nominal and Clean disables all noise.Source: §A.5 · p. 15

B.1 EVALUATION PROTOCOL

• Each deployment scenario is evaluated on all four LIBERO suites with 20 episodes per task and three seeds, and reported success rates are averaged over tasks and seeds.The seeds vary policy sampling noise while initial states are shared.Source: §B.1 · p. 15
• For each backbone, methods share a pretrained checkpoint, action horizon, per-task initial states, and scenario parameters.Hidden noise states are reset at the start of each episode.Source: §B.1 · p. 15
• Realized execution errors depend on each policy’s state trajectory because the deployment scenarios evolve in closed loop.Table A4 reports scenario settings using J1–J4 / J5–J7 values unless otherwise stated.Source: §B.1 · p. 16, Table A4 · p. 16

B.2 CONFIGURATION OF OUR METHOD

• A rank-4 LoRA adapts the frozen backbone from residuals on executed steps, updating every eight chunks once a 32-chunk feedback buffer is full.The policy executes five of each 50 predicted steps; simulation carries adaptation state across tasks in a suite but clears the buffer at task boundaries.Source: §B.2 · p. 16

B.3 ACTION AND OBSERVATION SPACE

• Actions encode end-effector translation, axis-angle rotation, and a binary gripper command; residual correction excludes the gripper and unused padding dimensions.The observation combines proprioception, two RGB views, and the language instruction.Source: §B.3 · pp. 16–17

B.4 LORA AND FLOW-MATCHING OPTIMIZATION

• The rank-4 LoRA adds 2,764,800 trainable parameters, or 0.085% of the π0 backbone’s 3.24 B parameters, while the vision-language backbone remains frozen.The policy uses a flow-matching loss and Adam optimization.Source: §B.4 · p. 17

B.5 COMPUTATIONAL SETUP

• The simulation control loop runs at 20 Hz while physics integrates at 500 Hz; one LoRA Adam step takes about 50 ms.Multi-step flow-matching inference takes about 80 ms per chunk.Source: §B.5 · p. 17

B.6 REAL-ROBOT SETUP

• On the physical Piper arms, the policy uses joint-target actions and computes residuals from measured joint state against the uncorrected policy target.For DOB+Ours, residuals are measured before disturbance-observer correction, so adaptation targets error the DOB does not cancel rather than its correction.Source: §B.6 · pp. 17–18
• Real-robot updates are scheduled every eight feedback chunks and applied to subsequent predictions; adaptation state persists across 25 evaluation episodes of a task.Those same evaluation episodes populate the feedback buffer, and updates begin when it is full.Source: §B.6 · p. 17
• For OOD evaluation, adaptation starts from π0.5 with initialized LoRA parameters, fresh optimizer state, and an empty buffer, then persists across OOD episodes and object variants.The residual and pseudo-target are converted from absolute joint coordinates to the action representation used in fine-tuning.Source: §B.6 · p. 18

C RESIDUAL MAGNITUDE ACROSS SCENARIOS

• 38.7%–56.3% residual magnitude across RoboStress scenarios, versus 23.4% in Clean; physical-arm values are 32.9% for Robot A and 35.0% for Robot B.The simulation residuals use normalized end-effector action units at 20 Hz, while physical-robot residuals use joint-space radians at 30 Hz, so the two groups are not on a common scale.Source: Table C5 · p. 18, C RESIDUAL MAGNITUDE ACROSS SCENARIOS · p. 18

D.1 IMPACT OF THE ANCHOR LOSS

• Removing the anchor term lowers average success across the four LIBERO suites from 44.2% to 41.8%, with decreases in all seven deployment scenarios.The largest decrease is 4.6 points under Heavy Payload; Aged Transmission and Aged Joint-Shoulder each decrease by 3.7 points.Source: Table D6 · p. 19, §D.1 · p. 19

D.2 PERFORMANCE ACROSS SEVERITY LEVELS

• Each scenario is tested at five severity levels, L1 to L5; L5 uses the full Table 2 setting, with thermal ramps and worn-joint settings handled as specified.Source: §D.2 · p. 19, Table D7 · p. 19
• 65.2% to 60.5% for Thermal Drift-Stribeck and 92.4% to 84.8% for Thermal Drift-Backlash are the smallest reported drops.Source: §D.2 · p. 19
• Aged Joint-Elbow stays around 90% success at every severity level.Source: §D.2 · p. 20
• LIBERO-Long is most affected across suites, falling to about two-thirds of its L1 success rate as errors accumulate over its longer horizon.Source: §D.2 · p. 20
• Self-compensating VLA generally matches or beats the base policy; its gap widens with severity on Heavy Payload and Aged Joint, but stays roughly constant on Thermal Drift.Source: §D.2 · p. 20
• 84.3% to 56.8%: self-compensating VLA’s average success across four suites falls most under Heavy Payload from L1 to L5.Source: §D.2 · p. 19, Figure D1 · p. 21