Reproduce “Sensing Which Modality Matters: Evidence-Gated Regularization for Robust VLA Policies” 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
Yue Yang, Diego Romeres, Chiori Hori, Gedas Bertasius, Daniel Szafir, Siddarth Jain
Researchers from the University of North Carolina at Chapel Hill and Mitsubishi Electric Research Laboratories (MERL) developed Evidence-Gated Regularization (EGR) to enhance Vision-Language-Action (VLA) policy robustness by mitigating "modality entanglement," where policies develop spurious dependencies on irrelevant sensory inputs. The method leverages explicit, task-grounded evidence scores to gate consistency losses during training, leading to improved task success rates by up to 190% on real-robot tasks under sensor corruption and novel physical distractors.
Key insights
- Modality entanglement, including nuisance sensitivity and single-modality insufficiency, significantly degrades VLA policy performance when faced with sensor corruptions or distractors.
- Explicitly grounding sensor relevance in task structure, rather than relying on brittle learned correlations, offers a robust way to guide multimodal policy learning.
- An evidence-gated regularization approach, applied during training, can effectively reduce spurious sensor dependencies and improve generalization without adding runtime complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.