mlsanityData to 7 Oct 2026

Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models

arXiv 2609.39820 ↗First posted 30 Sep 2026cs.RO cs.AIcited by 0

On the map: Vision-language-action models → Failure recovery and robustness

Abstract (arXiv)

Vision-language-action (VLA) models generalize broadly across robotic manipulation tasks, but complex environments require balancing task success with unintended contact. Runtime shields can correct individual actions, but they leave the underlying policy unchanged, so repeated disagreements may create a persistent policy-shield mismatch that blocks task progress. To address this challenge, we introduce FailBank, a four-stage self-evolving framework that converts runtime feedback into persistent policy improvement. During collection, a fixed CBF-based safety module serves as an observe-only teacher, producing counterfactual corrections while the policy remains in control. Outcome-aware admission then converts useful proposals into corrective targets and retains successful uncorrected actions as quiet anchors for guarded LoRA updates. We evaluate FailBank on the VLA-Arena benchmark across two difficulty levels and two VLA backbones. Compared with the base policies, FailBank improves the joint success-cost operating point. Across the two backbones, FailBank improves task success rate by 8.5 and 6.9 percentage points, while reducing policy-induced cumulative cost by 35.6% and 23.8%, respectively. Compared with runtime shielding, FailBank raises task success rate by 25.4 and 9.5 percentage points, while maintaining comparable policy-induced cumulative cost. These results show that runtime feedback can serve as persistent policy supervision rather than only as a temporary action constraint.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Failure recovery and robustness, one of 13 directions in Vision-language-action models. Builds on 28 of its 32 references that are in the map. First posted on arXiv 30 Sep 2026.

Builds on

28 of 32 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models27 Dec
  2. 2025SAFE: Multitask Failure Detection for Vision-Language-Action Models11 Jun

Other references · 26

  1. 2023RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Controlcited by 2,993
  2. 2023Diffusion Policy: Visuomotor Policy Learning via Action Diffusioncited by 2,730
  3. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,297
  4. 2025π0.5: a Vision-Language-Action Model with Open-World Generalizationcited by 1,653
  5. 2023Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardwarecited by 1,490
  6. 2023LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learningcited by 1,338
  7. 2017Constrained Policy Optimizationcited by 928
  8. 2025FAST: Efficient Action Tokenization for Vision-Language-Action Modelscited by 643
  9. 2021CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Taskscited by 564
  10. 2025SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Roboticscited by 507
  11. 2024Evaluating Real-World Robot Manipulation Policies in Simulationcited by 484
  12. 2024RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robotscited by 391
14 more
  1. 2025UniVLA: Learning to Act Anywhere with Task-centric Latent Actionscited by 386
  2. 2019Learning by Cheatingcited by 361
  3. 2016Control Barrier Function Based Quadratic Programs for Safety Critical Systemscited by 261
  4. 2018HG-DAgger: Interactive Imitation Learning with Human Expertscited by 245
  5. 2020Recovery RL: Safe Reinforcement Learning with Learned Recovery Zonescited by 156
  6. 2018Residual Policy Learningcited by 146
  7. 2020Human-in-the-Loop Imitation Learning using Remote Teleoperationcited by 103
  8. 2025SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learningcited by 69
  9. 2022Efficient Learning of Safe Driving Policy via Human-AI Copilot Optimizationcited by 44
  10. 2023Model-Based Runtime Monitoring with Interactive Imitation Learningcited by 37
  11. 2025VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layercited by 32
  12. 2022Provably Safe Deep Reinforcement Learning for Robotic Manipulation in Human Environmentscited by 16
  13. 2026Neuro-Symbolic Safety Guidance for Vision-Language-Action Models via Constrained Flow Matchingcited by 4
  14. 2026ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Modelscited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Benchmarking Vision-Language-Action Models on SO-101: Failure and Recovery Analysis
  2. 2026Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models
  3. 2026SEES: A Self-Evolving Embodied System via Failure-Guided VLA Policy Adaptation
  4. 2026RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models
  5. 2026CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies
  6. 2026PhysReflect-VLA: Physical Feasibility and Self-Reflective Regulation for Reliable Vision-Language-Action Policies
  7. 2026Find Something You Can't Do: Agentic Real-World Reinforcement Learning for Self-Improving VLA Models
  8. 2025EVOLVE-VLA: Test-Time Training from Environment Feedback for Vision-Language-Action Models