Abstract:

Vision-Language-Action (VLA) policies integrate multiple sensors to support robust robotic task execution, yet limited and homogeneous demonstrations can encourage reliance on spurious inter-sensor correlations rather than task-relevant evidence. We call this failure modality entanglement. It takes two forms: nuisance sensitivity, where corrupting an uninformative sensor still degrades the policy, and single-modality insufficiency, where the policy cannot act on the one informative sensor that remains intact. We study both failures on a BEHAVIOR-1K benchmark of 47 skills and on two real robots, and address them with Evidence-Gated Regularization (EGR), a training objective that leverages simple task knowledge to guide the policy toward the most relevant sensor at each frame. Our study shows that with limited demonstrations more sensors can make a policy more fragile, that clean-input success hides both failures, and that gating the training loss by task-relevant evidence reduces them without new data or architectural changes. EGR more than doubles success under single-sensor fallback in simulation and raises success under physical distractors from 30% to 85% on a bi-manual robot.