G^3VLA: Geometric Inductive Bias for Vision-Language-Action Models

G^3VLA. Geometric Inductive Bias for Vision-Language-Action Models. By Yue Peng and seven others, from New York University Shanghai, the Technical University of Denmark, and two other institutions. Published on the fourteenth of October. No code repository is linked to it.

Vision-language-action models have pushed generalist robot manipulation forward by borrowing the semantic knowledge of pretrained vision-language backbones. Their visual representations, though, are usually grounded in 2D image coordinates, not in calibrated camera geometry. The gap is sharpest on robots with several cameras. There, the views are coupled by known intrinsics and extrinsics: the system knows exactly how each camera sits relative to the others. Yet each view is still processed as an independent image, as if that knowledge did not exist.

G^3VLA is a camera-aware geometric module that puts that calibrated structure back in. It injects it into the stream of visual tokens inside a pretrained VLA, and it does so without altering the model's action space or its original action-learning objective. The policy keeps predicting actions the way it always did. What changes is that the tokens it sees now carry where they sit in camera geometry.

Three components establish geometrically consistent representations across views. Ray embeddings conditioned on each camera's intrinsics. A projective positional encoding, called PRoPE. And bidirectional fusion across camera views. Training the module needs geometric supervision, and that comes in two forms. One is ground-truth point maps. The other is predictions from a teacher model, gated by the teacher's confidence. Either way, no depth sensors and no manual geometric annotations are needed.

Built on top of pi zero, G^3VLA consistently improves performance on LIBERO and RoboCasa24, under both supervision regimes. With ground-truth supervision, it also improves the five-task averages on RoboTwin2.0 with three cameras. Experiments on a camera-shifted version of LIBERO-Goal, and on real robots, show improved robustness to camera viewpoints the policy never saw.

Controlled calibration studies, run on both pi zero and pi zero point five, check where the gains come from. They depend on correct camera geometry: the improvement is tied to the calibration actually being right. At the same time, the gains stay stable when the calibration is only mildly perturbed.

The authors present this as a simple and general mechanism. Ground the visual tokens of a pretrained robot policy in calibrated camera geometry, and you get better geometric consistency, better reasoning across views, and better robustness to viewpoint. The robot's cameras were always calibrated. The step here is letting the policy's tokens know it.