Original title: Sensing Which Modality Matters: Evidence-Gated Regularization for Robust VLA Policies
Yue Yang, Diego Romeres, Chiori Hori, Gedas Bertasius +2 · September 2026
Robots that use cameras and touch sensors often learn bad habits from limited training demonstrations. They latch onto accidental links between their sensors instead of truly understanding the task. Because of this, a robot might fail if a useless sensor gets blocked, or struggle if only one good sensor is left working. To fix this, researchers created Evidence-Gated RegularizationA training method that adjusts focus based on how useful each sensor is for the task.. This training method measures how relevant each sensor is in every moment. It trains the robot to ignore sensors that do not matter and rely completely on sensors that do. It adds no extra running time during use. In real tests with physical distractors, this method increased a two-arm robot's success rate from 30% to 85%. On a single-arm robot using touch sensors, success climbed from 55% to 70%.
You're making a recipe with fresh and spoiled ingredients. If you use all of them without checking, the dish tastes bad. In this paper, robots learn to test which sensors are reliable during training, so they only use the good ones when performing tasks.
In real environments, robot sensors often get blocked or fail, making robots unreliable. This method helps robots adapt when sensors are corrupted, making them safer and more practical for everyday use.
The abstract doesn't specify limitations beyond the tested setups. It validates on specific benchmarks and robot embodiments, but doesn't address performance on other tasks or sensor types.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Sensing Which Modality Matters: Evidence-Gated Regularization for Robust VLA Policies — Yue Yang, Diego Romeres, Chiori Hori et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.