mlsanityData to 14 Oct 2026

FailBench: How Reliable are VLMs at Judging Robot Task Success?

arXiv 2609.03611 ↗First posted 3 Sep 2026cs.RO cs.AIcited by 0

On the map: Multimodal LLMs / vision-language models → Embodied and 3D vision-language models

Abstract (arXiv)

Vision-Language Models (VLMs) are increasingly used to evaluate robot manipulation outcomes, but existing benchmarks offer limited evidence of cross-domain generalization. We introduce FailBench, a benchmark for robot failure detection comprising 2,197 manipulation attempts across 14 public sources (12 real-world, 2 simulated). In FailBench, 75% of failures occur naturally, and six real-world sources come from non-failure-detection datasets. Evaluating 13 VLM-based detectors, we find the best model achieves only 0.77 mean balanced accuracy. Notably, models fine-tuned for failure detection consistently underperform general-purpose VLMs and their own pretrained baselines. Performance depends heavily on required visual evidence: models approach saturation when outcomes depend on observable object motion, but degrade to near-chance (<0.60 balanced accuracy) on contact-intensive assembly tasks. Error analysis reveals a systematic bias toward predicting success under ambiguous evidence, which persists even with increased reasoning effort. Finally, we show that input-level intervention--spatially localizing and cropping outcome-relevant regions--improves the top detector by 2.4 percentage points without extra training.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Embodied and 3D vision-language models, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 47 of its 49 references that are in the map. First posted on arXiv 3 Sep 2026.

Builds on

47 of 49 references.

Closest ancestors · 6

Same direction, newest first

  1. 2026Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning12 Feb
  2. 2025Guardian: Detecting Robotic Planning and Execution Errors with Vision-Language Models1 Dec
  3. 2025Robobench: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models as Embodied Brain20 Oct
  4. 2025FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models2 Oct
  5. 2025I-FailSense: Towards General Robotic Failure Detection with Vision-Language Models19 Sep
  6. 2024Automating Robot Failure Recovery Using Vision-Language Models With Optimized Prompts6 Sep

Other references · 41

  1. 2024GPT-4o System Cardcited by 4,588
  2. 2025Qwen3-VL Technical Reportcited by 2,793
  3. 2024Evaluating Real-World Robot Manipulation Policies in Simulationcited by 495
  4. 2024LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methodscited by 465
  5. 2026Gemma 4 Technical Reportcited by 259
  6. 2023RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shotcited by 183
  7. 2025VLA-RL: Towards Masterful and General Robotic Manipulation with Scalable Reinforcement Learningcited by 176
  8. 2023REFLECT: Summarizing Robot Experiences for Failure Explanation and Correctioncited by 175
  9. 2024AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulationcited by 125
  10. 2024GRAPE: Generalizing Robot Policy via Preference Alignmentcited by 109
  11. 2025SAFE: Multitask Failure Detection for Vision-Language-Action Modelscited by 80
  12. 2024Unpacking Failure Modes of Generative Policies: Runtime Monitoring of Consistency and Progresscited by 69
29 more
  1. 2025WorldGym: World Model as An Environment for Policy Evaluationcited by 64
  2. 2025RoboArena: Distributed Real-World Evaluation of Generalist Robot Policiescited by 64
  3. 2024Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detectioncited by 63
  4. 2026Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisonscited by 60
  5. 2026RoboReward: General-Purpose Vision-Language Reward Models for Roboticscited by 53
  6. 2025Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policiescited by 52
  7. 2025AutoEval: Autonomous Evaluation of Generalist Robot Manipulation Policies in the Real Worldcited by 48
  8. 2025Failure Prediction at Runtime for Generative Robot Policiescited by 42
  9. 2025RoboFAC: A Comprehensive Framework for Robotic Failure Analysis and Correctioncited by 41
  10. 2024Autonomous Improvement of Instruction Following Skills via Foundation Modelscited by 32
  11. 2025RoboEval: Where Robotic Manipulation Meets Structured and Scalable Evaluationcited by 32
  12. 2025Are We on the Right Way to Assessing LLM-as-a-Judge?cited by 24
  13. 2025Reliable and Scalable Robot Policy Evaluation with Imperfect Simulatorscited by 14
  14. 2025Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbolscited by 14
  15. 2026Foresight: Failure Detection for Long-Horizon Robotic Manipulation with Action-Conditioned World Model Latentscited by 11
  16. 2026Learning Actionable Manipulation Recovery via Counterfactual Failure Synthesiscited by 10
  17. 2026PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditingcited by 9
  18. 2025Score the Steps, Not Just the Goal: VLM-Based Subgoal Evaluation for Robotic Manipulationcited by 8
  19. 2026Hy-Embodied-VLM-1.0: Efficient Physical-World Agentscited by 8
  20. 2026ManipulationNet: An Infrastructure for Benchmarking Real-World Robot Manipulation with Physical Skill Challenges and Embodied Multimodal Reasoningcited by 6
  21. 2026From Passive Observer to Active Critic: Reinforcement Learning Elicits Process Reasoning for Robotic Manipulationcited by 6
  22. 2026Failure Identification in Imitation Learning Via Statistical and Semantic Filteringcited by 6
  23. 2026Progress Reward Modeling for Robotic Learning: A Comprehensive Surveycited by 5
  24. 2026Eval-Actions: Fine-Grained Execution Quality Evaluation for Robotic Manipulationcited by 4
  25. 2026VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Modelscited by 3
  26. 2026PhAIL: A Real-Robot VLA Benchmark and Distributional Methodologycited by 2
  27. 2026SAFECAST: Robust Failure Detection for VLA Policies with Contrast-Set Training and Calibrationcited by 2
  28. 2026ArmnetBench v0.1: Parallel Real-World Evaluation of Manipulation Policies on a Low-Cost Arm Farmcited by 1
  29. 2026GUARD: Grounding Uncertainty and Ablation-Based Risk Detection for Diffusion-Based VLAscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025ManipBench: Benchmarking Vision-Language Models for Low-Level Robot Manipulation
  2. 2026KITE: Keyframe-Indexed Tokenized Evidence for VLM-Based Robot Failure Analysis
  3. 2024Addressing Failures in Robotics using Vision-Based Language Models (VLMs) and Behavior Trees (BT)
  4. 2026GroundBench: A Factorized, Counterfactual Benchmark for Locating VLM Affordance Failures
  5. 2026FailureSpot: Label-Efficient Timestamp-Level Failure Detection for Vision-Language-Action Models
  6. 2026RAFAIL: Relationship-Aware Failure Detection for Robotic Manipulation ↗
  7. 2026Benchmarking Vision-Language-Action Models on SO-101: Failure and Recovery Analysis
  8. 2026Perturbation-Based Epistemic Uncertainty for Failure Detection in Vision-Language-Action Models