VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation
VersaCamVLA creates a camera-configurable Vision-Language-Action (VLA) system that handles any number of camera views. It learns a unified scene-token interface to compress variable sets of posed RGB views into fixed latent tokens. These tokens are then fed into a base VLA as supplementary visual information, allowing the robot to perform robust manipulation regardless of the actual camera setup during deployment.