Lighting-Robust Behavioral Cloning via Bayesian Visual Encoding for Robot Manipulation

Abstract

Behavioral cloning (BC) has demonstrated strong performance in robot manipulation by learning visuomotor policies directly from expert demonstrations. However, under illumination-induced distribution shifts, vision-based BC policies often become sensitive to spurious visual variations, leading to degraded reliability and generalization. To address this issue, we propose a lighting-robust BC framework that incorporates Bayesian deep learning into the visual encoder. Specifically, we model a posterior distribution over convolutional weights and optimize the evidence lower bound, where a Kullback–Leibler regularization term constrains feature learning under illumination variability. In simulation, the proposed method achieves higher success rates and shorter rollout lengths than deterministic baselines, while maintaining strong sample efficiency with a single Monte Carlo sample. In real-world experiments across seven illumination regimes, including distribution-shifted conditions, our method consistently outperforms BC baselines and remains stable under severe lighting changes. These results indicate that Bayesian regularization provides an effective inductive bias for learning visual representations that improve robustness and generalization in vision-based robot manipulation.