G^3VLA adds calibrated camera geometry to pretrained VLA models
An arXiv paper proposes G^3VLA, a camera-aware geometric module that injects calibrated camera geometry into the visual-token stream of a pretrained VLA model without altering its action space or original action-learning objective. It combines intrinsic-conditioned ray embeddings, projective positional encoding (PRoPE) and bidirectional cross-view fusion, with geometric supervision from ground-truth point maps or confidence-gated π³X teacher predictions, so no depth sensors or manual geometric annotations are needed. The authors report consistent gains on LIBERO and RoboCasa24 with π0, improved five-task RoboTwin2.0 averages with three cameras under ground-truth supervision, and better robustness to unseen camera viewpoints on camera-shifted LIBERO-Goal and real robots.
If the results replicate, this way of injecting geometry without touching the action space offers a fairly general route to better multi-camera robot policies.