D-Robotics and University of Tokyo Propose RawVLA for Improved Lighting Robustness

The research team from D-Robotics, the University of Tokyo, and other institutions introduced RawVLA, which integrates a learnable imaging module ahead of a frozen VLA policy. Operating directly on RAW sensor data, the module adjusts imaging parameters based on the current frame and historical information to produce images better suited for downstream manipulation tasks. During training, the VLA policy remains frozen, allowing the imaging module to be optimized independently. This approach aims to make robot manipulation more robust to lighting variations, a common challenge in real-world deployment. By adapting the imaging process at the sensor level, RawVLA enhances the VLA model's illumination robustness without retraining the policy. The method demonstrates a promising direction for improving embodied AI systems' reliability under diverse lighting conditions, and highlights the potential of sensor-level adaptation in robotics.