PaperlayerView exact revision on arXiv

Source-linked AI summary

Training-Free Action Correction for VLA Model Failures via Language Feedback

Owen Kwon, Pablo Ortega-Kral, Arthur Bucker, Jean Oh

arXiv:2608.29967v1cs.ROcs.AI

TL;DR

VLA deployment failures are poorly understood, particularly whether systematic errors can be corrected without retraining. CorrectVLA converts task-level language into additive action adjustments while keeping policy weights frozen, and it recovers execution-misalignment failures in simulation and real-robot deployment. Its supported boundary is that correction works when strategic correctness is preserved, whereas semantic comprehension failures require other interventions.

  • Problem

    Deployment adaptation lacks evidence about when VLA failures can be corrected without retraining, despite systematic failures under changed environments.

  • Method

    CorrectVLA uses a single task-level natural-language correction, grounded into temporal action adjustments and reused across rollouts without updating policy weights.

  • Results

    95% success (19/20) after environment shift versus 10% (2/20) for the base policy, while simulation recovers execution-misalignment failures across in-distribution and OOD tasks.

  • Takeaways & Limitations

    Inference-time correction is useful for execution misalignment when policies retain strategic correctness, establishing a practical boundary for deployment adaptation.

  • Takeaways & Limitations

    Predefined magnitude ranges can require multiple correction rounds, and a single task-level correction may not generalize when failures vary across episodes.

Abstract

from arXiv · show

Vision-Language-Action (VLA) models demonstrate strong semantic understanding yet exhibit systematic failures during deployment. The conditions under which these failures occur, and whether they can be corrected without retraining, remain poorly understood. In this paper, we take steps toward addressing this gap. We present CorrectVLA, a framework that translates task-level natural language corrections into additive action magnitude adjustments without modifying policy weights. A human provides a single task-level correction, applied uniformly across all rollouts without per-episode intervention. In simulation, CorrectVLA recovers execution misalignment failures across both in-distribution and OOD tasks. In real-robot experiments on a UFactory xArm7 under environment shift, CorrectVLA restores near-perfect success where the base policy almost entirely breaks down, generalizing across object locations and identities. Through a taxonomy of failure modes on LIBERO-90, we find that execution misalignment failures, where the policy reaches the correct target but miscalibrates action magnitudes, represent the correctable subset, while other failure modes where semantic comprehension itself breaks down are not amenable to this approach. The approach succeeds when policies possess strategic correctness and fails when fundamental comprehension is absent, establishing a practical operational boundary for inference-time correction.

I. INTRODUCTION

CorrectVLA addresses deployment failures by converting sparse human language feedback into additive action corrections while preserving frozen VLA policies. The approach targets execution misalignment and demonstrates recovery in simulation and under real-robot environment shift.

  • Deployment shifts can preserve semantic understanding while causing systematic action-magnitude errors or more fundamental failures involving target and goal comprehension.
  • Execution misalignment is the correctable failure subset, whereas semantic-comprehension failures are not amenable to inference-time correction.The distinction separates calibration errors from failures involving misunderstanding or absent comprehension.
  • CorrectVLA maps natural-language feedback to structured, per-dimension action adjustments without modifying VLA weights.The framework grounds sparse linguistic feedback into temporal corrections applied to frozen policy outputs.
  • A single task-level correction is reused across rollouts, enabling autonomous completion without per-episode intervention or policy updates.CorrectVLA uses sparse corrections rather than retraining or continuous human guidance.

II. RELATED WORK

Generalist VLA models achieve broad cross-task manipulation capabilities, but deployment shifts expose systematic action-magnitude and comprehension failures. These vulnerabilities motivate methods that characterize failures and identify appropriate interventions.

  • Large-scale pretraining gives generalist VLAs strong cross-task generalization and semantic capabilities across diverse manipulation tasks.
  • New lighting, viewpoints, or surfaces can produce systematic action-magnitude miscalibration during deployment.
  • Fine-tuning deployment adaptations require new demonstrations, GPU training, and substantial infrastructure.
  • A central gap is systematic understanding of why policies fail and which intervention fits each failure mode.

B. Runtime Policy Correction and Steering

Prior test-time correction methods require continuous supervision, multiple sampled trajectories, or policy retraining. CorrectVLA instead applies sparse retrospective language corrections as additive action-magnitude adjustments while keeping policy weights frozen.

  • Prior approaches: ITPS requires continuous human guidance during execution, while FOREWARN incurs real-time computational overhead by generating multiple trajectory samples.ITPS learns task feature weights through online gradient descent; FOREWARN selects among sampled candidates using world models and VLM verifiers.
  • Limitations of prior approaches: Sampling-based methods remain within the learned output distribution, and verification systems do not correct systematic action-magnitude errors.This motivates a correction mechanism that can alter action magnitudes directly rather than only select or verify sampled behaviors.
  • CorrectVLA: CorrectVLA applies additive action-magnitude corrections to generate behaviors outside the original distribution without changing policy weights.The method provides a complementary alternative to sampling-based adaptation and verification-only systems.
  • Prior approaches: Comparative language-feedback methods require multiple execution rounds and pairwise comparisons, whereas shared autonomy adapts during execution from inferred user intentions.These approaches differ in whether feedback is comparative, continuous, or integrated into blended control.
  • CorrectVLA: CorrectVLA uses a sparse retrospective language correction after observing a complete failure, requiring minimal per-task review without fine-tuning or continuous guidance.The correction is applied after failure rather than supplied continuously during execution.

A. Problem Statement

The problem is to construct sparse, deployable corrections that convert near-miss VLA trajectories into successful executions while preserving policy weights. CorrectVLA focuses on execution misalignment, which preserves strategic correctness but miscalibrates action magnitudes.

  • A. Problem Statement: The correction function adds per-timestep, per-dimension adjustments to frozen policy actions to seek successful task completion.The formulation requires sparsity, policy preservation, and deployment without new demonstrations or GPU training.
  • A. Problem Statement: The correction must use few human feedback instances, keep policy weights frozen, and avoid new demonstrations or GPU training.
  • A. Problem Statement: Across 133 LIBERO-90 OOD evaluations, Task Misunderstanding dominates, while Perception Failure, Execution Misalignment, and Multi-step Planning occur less often.
  • A. Problem Statement: Execution Misalignment is the primary correction target because strategic correctness remains intact despite miscalibrated action magnitudes.
  • A. Problem Statement: CorrectVLA assumes that a small number of magnitude adjustments can fix a near-miss and that a human can identify the critical error region.

IV. METHODS

CorrectVLA converts one task-level natural-language correction into a structured, time-localized action adjustment while preserving the policy’s high-level behavior.

  • Phase A: Structured Parameter Extraction: Phase A parses feedback into action dimension, direction, magnitude term, and temporal window.For example, “move left more in 1–2 seconds” becomes a structured correction.
  • Phase C: Temporal Distribution: Phase C distributes each correction across its specified time window with a piecewise-linear envelope that peaks at the midpoint.The correction ramps from zero to peak magnitude and returns to zero outside the window.
  • Phase C: Temporal Distribution: A single “down little, 2.0–2.3s” correction redirects the gripper to recover a missed bowl grasp.The example preserves the policy’s high-level structure and adjusts only the critical action magnitude.

D. Complete Grounding Function

The grounding function adds localized correction terms to the frozen policy action, allowing sparse feedback to combine across dimensions and persist across rollouts.

  • Complete Grounding Function: Multiple feedback instances are aggregated across action dimensions before producing the corrected action.Corrections are sparse in feedback space, localized in time, and superpose additively.
  • Complete Grounding Function: The corrected action is formed by adding a correction term δ_d(t) to the policy action a_t,d.This additive form changes action magnitudes without modifying policy weights.
  • Complete Grounding Function: Policy weights remain frozen, and task-level feedback is reused across all rollouts without per-episode human intervention.The design separates one-time task correction from repeated policy execution.

V. EXPERIMENTAL SETUP

The experiments compare frozen-policy correction against an uncorrected base policy and an autonomous LLM baseline across simulation and shifted real-robot settings.

  • Benchmark and Models: Simulation uses LIBERO with Franka Panda and evaluates OpenVLA-OFT and π0.5 as base policies.OpenVLA-OFT is a 7B-parameter VLA fine-tuned on 40 LIBERO tasks; π0.5 is evaluated in simulation and on the real robot.
  • Benchmark and Models: The real-robot setup relocates the xArm7 base, shifting camera viewpoint and workspace geometry while keeping the objects the same.The xArm7 uses a Robotiq 2F-85 gripper and Intel RealSense camera.
  • Baselines: The Base Policy is the frozen policy without correction, while the LLM Baseline autonomously infers displacement and bias from failure videos.CorrectVLA instead uses one human task-level correction formatted as JSON.
  • Baselines: After the base shift, π0.5 fails on both pick-and-place tasks, whereas CorrectVLA restores success using one task-level correction.The correction requires neither per-episode intervention nor retraining.
  • Evaluation Scope: CorrectVLA is evaluated on simulation results spanning in-distribution and out-of-distribution tasks, with corrections applied only to execution-misalignment cases.This setup tests the method on the failure category identified as correctable.

C. Evaluation Protocol

The evaluation tests task-level corrections across simulation and three real-robot generalization conditions after changing the robot’s environment or task instance.

  • Simulation: Simulation uses 50 trials per task on 40 in-distribution tasks and 5 trials per task across all 90 LIBERO-90 tasks for OOD evaluation.Corrections are applied only where the base policy shows execution misalignment.
  • Real Robot: Real-robot evaluation covers same object with changed environment, same object at a different location, and different object at a different location.These conditions test whether one task-level correction transfers across rollout variations.
  • Real Robot: Table II evaluates whether one task-level correction recovers performance across all three conditions after the robot base shift.The protocol focuses on pick-and-place with a can under shifted deployment conditions.
  • Real Robot: π0.5 achieves 95% success over 20 trials in the original setting before the robot base is shifted.The reported original-environment result is 19/20 successes.

D. Action Space Adaptation

CorrectVLA adapts Cartesian-space human corrections to the xArm7’s joint-velocity action space while targeting failures that preserve strategic correctness. On LIBERO-90, execution misalignment is the correction-amenable failure mode, and task-level corrections improve in-distribution recovery.

  • Action-space conversion: Cartesian corrections are converted into joint velocities through the Jacobian pseudoinverse for the xArm7’s joint-velocity controller.The correction is specified in intuitive directions such as “move left” or “move up.”
  • Correctable failure mode: Execution Misalignment preserves strategic correctness while miscalibrating action magnitudes, making it amenable to inference-time correction.The remaining taxonomy modes require training-level interventions.
  • In-distribution recovery: CorrectVLA recovers 65 of 139 in-distribution failures, raising π0.5 success from 93.0% (1861/2000) to 96.3% (1926/2000).The task-level corrections are applied uniformly across rollouts, while the LLM baseline recovers nothing.

2) In-Distribution Recovery:

CorrectVLA recovers execution-misalignment failures across OOD simulation and real-robot deployment, including environment and object shifts. Its gains depend on preserving task understanding rather than correcting semantic failures.

  • OOD recovery: 33 execution-misalignment failures are recovered on LIBERO-90, raising success from 27.1% (122/450) to 34.4% (155/450).The recovery rate is 10.1% (33/328), and task misunderstanding failures are not recovered.
  • Real-robot recovery: 95% real-robot success (19/20) is restored after relocation, versus 10% (2/20) for the shifted base policy.Success reaches 100% for same-location (10/10) and different-location (5/5) conditions, and 80% for different object and location (4/5).
  • Failure boundaries: Task Misunderstanding and Perception Failure comprise over 70% of failures across both models, whereas Execution Misalignment is the primary correction candidate.Magnitude adjustment cannot compensate for absent comprehension.
  • Sparse correction: A small number of targeted adjustments at critical trajectory moments can redirect failed executions while preserving the policy’s high-level structure.Examples use one correction for “pick up bowl on cabinet” and two corrections for “open middle drawer.”
  • Generalization: Task-level corrections generalize across rollout variations and real-robot conditions without per-episode intervention or retraining.The demonstrated conditions include same-location, different-location, and different-object shifts.

VIII. LIMITATIONS AND FUTURE WORK

The paper frames CorrectVLA as a targeted alternative to retraining for execution-magnitude errors, while identifying semantic failures and several deployment assumptions as boundaries. Future work targets adaptive correction magnitudes and broader failure modes.

  • Limitations: Predefined language-to-magnitude ranges can make applied corrections vary from intended corrections and require multiple rounds.The paper maps “slightly,” “more,” and “much” to predefined magnitude intervals.
  • Limitations: A single task-level correction may not generalize when failure modes vary substantially across episodes.The method assumes one representative failure captures the dominant task misalignment.
  • Future work: Future work proposes learning magnitude estimates from accumulated correction failures and using object-centric, randomized, and spatial-reasoning methods for broader failure modes.These directions target task misunderstanding, perception failure, and planning failures beyond magnitude adjustment.
  • Conclusion: CorrectVLA addresses execution misalignment without modifying policy weights, using task-level corrections across simulation and real-robot shifts.Real-robot success reaches 95% where the base policy drops to 10%.
  • Operational boundary: Inference-time correction succeeds when strategic correctness is preserved and fails when fundamental comprehension is absent.The paper presents this distinction as a practical deployment principle for selecting interventions.
Loading 2608.29967v1…