The practical deployment of robots is often hampered by the discrepancy between controlled training environments and the dynamic, unpredictable nature of real-world scenarios. Specifically, camera repositioning and remounting are common, yet existing Vision-Language-Action (VLA) policies falter when camera extrinsics aren't explicitly provided, leading to fragile performance. This limitation is particularly acute in tasks demanding robust visual perception. The researchers behind CamVLA propose a paradigm shift: the policy itself should infer camera geometry rather than relying on external, often unavailable, calibration data.
Camera-Centric Action Generation for Robustness
CamVLA introduces a novel approach that decouples manipulation controls from static camera geometry. Instead of outputting actions in a fixed robot base frame, it predicts a camera-centric end-effector action, effectively defining movements relative to the camera's local frame. This is complemented by predicting a 6-DoF hand-eye matrix, which establishes the relationship between the camera and the robot base. A deterministic geometric transformation then fuses these two predictions to generate a robot base-frame action. This disentanglement allows for pose-independent action generation within the camera's view, while simultaneously grounding it in the physical world through geometric reasoning. This core innovation dramatically enhances the robustness of CamVLA robot control in diverse, unseen viewpoints.