Research map / Vision-language and multimodal / Robot action and navigation

Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AFIL, which trains a dual action generator on success and failure rollouts from a pretrained pi0.5 VLA and steers diffusion/flow sampling away from failure actions with a cosine-distance-adaptive negative guidance.

Problem

VLA policies trained only on success demonstrations are brittle: small execution errors compound into unrecoverable out-of-distribution failures, and existing failure-learning methods rely on handcrafted failure modes or human intervention.

Approach

Online-generated failure rollouts from a pretrained VLA train a second action head sharing the VLM backbone; at sampling time, negative-prompting-style guidance subtracts the failure score with strength proportional to 1 minus the cosine similarity between success and failure predictions at each diffusion step.

Key results

• On LIBERO, DAG-VLA with AFIL raises pi0.5 average success from 96.9% to 98.4% (LIBERO-10: 92.4% to 95.8%).
• On ManiSkill in-domain long-horizon tasks, Adaptive-FIL reaches 56.7% and 71.3% versus 49.3% and 59.3% for Static-FIL.
• On out-of-domain tasks, AFIL achieves 62.7 / 48.0 / 64.0 versus 53.3 / 41.3 / 54.7 for pi0.5 with failure correction.

Surprising findings

• Static negative guidance with fixed strength underperforms adaptive guidance, and larger guidance scales (alpha 2.0 and 5.0) reduce success rates.

Limitations

• Evaluation limited to ManiSkill and LIBERO simulation tasks; no real-robot results in the main text.
• Guidance strength alpha and static lambda are tuned per setting and the paper's ablation covers only two tasks.
• Failure data depends on rollouts from a pretrained VLA, so failure distribution is tied to that policy.

Experimental setup

Datasets: ManiSkill LIBERO

Benchmarks: ManiSkill LIBERO

Metrics: success rate

Baselines: Success-only pi0.5 Success + failure correction Static-FIL Diffusion Policy Octo OpenVLA CoT-VLA

Builds on: pi0.5 negative prompting classifier-free guidance OpenVLA Diffusion Policy

Closest papers by meaning

• AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation (ICLR 2025 · similarity 0.93)
• RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• SAFE: Multitask Failure Detection for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.92)
• Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning (ICLR 2026 · similarity 0.89)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.89)