🔑 CamVLA: From 🔒 Fixed to Free Cameras 🗽
Current VLA policies are trained with a fixed camera and implicitly memorize the hand-eye coordinate mapping in their weights. When that camera is bumped, remounted, or hand-held at deployment, the mapping breaks. A mere 15° shift can crash success rates from 65.3% to just 6.3%. Existing view-robust methods try to fix this by feeding the policy known camera extrinsics, but accurate calibration is exactly what breaks under real-world camera changes.
Our key insight: the policy should not be told where the camera is, it should figure it out by itself. 📷 CamVLA uses an Action Head that answers "how should I move?" in the camera's own frame, and a Geometric Head that answers "where am I looking from?" by estimating the hand-eye pose from a single RGB image. A deterministic geometric transformation maps the action into the robot base frame, no calibration needed.
