RoHIL: Robust Human-in-the-Loop Robotic Reinforcement Learning Against Illumination Variations

Researchers introduced RoHIL, an offline fine-tuning framework for human-in-the-loop robot policies under lighting changes.

RoHIL relights previously collected trajectories with a world model, combines adaptation and retention data, and uses a Bellman-style regularizer to limit forgetting.

Across four real-robot manipulation tasks, it improved robustness when moving between workstations with different illumination.

The approach targets a common deployment problem: policies that work in one lighting setup often fail in another, and retraining from scratch is expensive.