ROBBRIEF
Robot policies trained on small demo sets learn wrong cross sensor wiring and freeze on clutter, yet a training fix, Evidence Gated Regularization, lifts success from 30% to 85% on cluttered setups.
Cover one of a robot's cameras, add some clutter to the table, and watch a state-of-the-art robot policy freeze, not because the model is weak, but because it learned the wrong cross-sensor wiring from a small set of demonstrations. A new arXiv preprint names that failure mode "modality entanglement" and ships a training-time fix that costs nothing at inference.
The method, Evidence-Gated Regularization (EGR), gates two consistency goals by a per-frame, per-sensor task-relevance signal: invariance on low-evidence sensors, single-sensor sufficiency on high-evidence ones. On the BEHAVIOR-1K benchmark, a StanfordVL suite of 1,000 everyday activities, 47 rollout-based skills show EGR lifts simulated success under uninformative-sensor corruption from 9.4% to 16.5%.
On real robots, the gap is larger. A bi-manual setup with two Kinova arms and three RGB cameras jumped from 30% to 85% success on tasks with physical distractors. A second setup pairing vision with GelSight tactile sensors climbed from 55% to 70%.
Under single-sensor fallback, the same method nudges success from 2.8% to 6.1%, a 3.3-point absolute change that is meaningful but far from a robust policy. The paper is a preprint, not peer-reviewed, and the gains are on the authors' own benchmark and two lab setups. Replication outside that benchmark is the next test.