ICML 2026 NeurVLA - Heungwoo/research GitHub Wiki

NeurVLA: Unleashing Failure-Handling Capability of VLA Models via Neural-Symbolic Reasoning — joint failure correction and prevention

Venue: ICML 2026 (Poster) Category: Reasoning Affiliations: Xuqi Liu, Minghe Gao, Juncheng Li, Siliang Tang

Problem

Vision-Language-Action (VLA) models are increasingly deployed on real robots, but execution failures frequently derail them on diverse, open-ended embodied tasks. The authors pinpoint two specific weaknesses in current systems: coarse-grained failure correction (recovery actions that are too blunt to fix the actual error) and unreliable failure prevention (no dependable mechanism to anticipate and avoid failures before they occur).

Method

NeurVLA is a neural-symbolic framework that jointly addresses failure correction and failure prevention via reasoning, and then internalizes these failure-handling capabilities directly into the VLA model rather than relying on an external module at inference time. The neural component supplies perception and action generation, while the symbolic-reasoning component handles structured failure analysis — diagnosing what went wrong and what should be done — covering both correcting a failure that has already occurred and preventing one that is about to. By distilling this reasoning back into the VLA policy, the model gains robust failure-handling behavior across diverse tasks. (Full architectural and quantitative details are in the paper's supplementary materials, which were not available at the time of writing.)

flowchart LR
    O[Vision + Language observation] --> N[Neural VLA perception/action]
    N --> S[Symbolic reasoning: failure correction + prevention]
    S --> I[Internalize failure-handling into VLA]
    I --> A[Robust action]
    A -. failure feedback .-> S

Results

The abstract reports that NeurVLA achieves "strong performance and robust generalization across diverse tasks." No specific numeric results are stated in the available abstract; quantitative benchmarks are referenced as being in the paper and supplementary materials.

Significance

Failure handling remains a major gap for VLA models in open-ended deployment. By unifying correction and prevention under one neural-symbolic reasoning framework and baking the resulting behavior into the policy itself, NeurVLA targets reliability — a prerequisite for trustworthy real-world robotic manipulation — without depending on a heavyweight external reasoning loop at run time.