Reproduce “Uncovering Vulnerability of Vision-Language-Action Models under Joint-Level Physical Faults” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Minsoo Jo, Taeju Kwon, Junha Chun, Youngjoon Jeong, Taesup Kim

Researchers at Seoul National University investigated the vulnerability of Vision-Language-Action (VLA) models to embodiment-side physical faults in robotic systems, such as joint locks or increased friction. They developed J-PARC, a lightweight residual calibration framework, which enhances VLA policy robustness by adaptively correcting actions based on inferred joint-fault regimes, leading to improved task success rates in both simulations and real-world experiments, and demonstrating generalization to unseen fault types.

Key insights

  • VLA models exhibit significant and heterogeneous vulnerability to joint-level physical faults, with performance degradation often driven by execution mismatch even when tasks remain kinematically feasible (e.g., under increased friction).
  • Existing robustness techniques, including robustness to action noise and local kinematic corrections like Constrained Inverse Kinematics (CIK), are ineffective against embodiment-side physical faults because they do not address the altered underlying action realization mechanism or cumulative trajectory drift.
  • A modular, residual calibration approach can successfully infer latent joint-fault regimes from robot dynamics and adaptively correct VLA policy actions, thereby maintaining task success without modifying or retraining the large base VLA model.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.