G^3VLA: Geometric Inductive Bias for Vision-Language-Action Models
Calibrated camera geometry improved robot task performance and robustness to unseen viewpoints in the authors’ tests.
G³VLA adds camera-aware geometry and cross-view fusion to pretrained vision-language-action models without changing their action space or training objective. On π₀, it improved LIBERO and RoboCasa24 performance using either ground-truth geometry or teacher predictions; the latter requires no depth sensors or manual geometric annotations. Camera-shifted simulations and real-robot experiments showed improved viewpoint robustness. Calibration tests found that the gains depend on correct camera geometry but tolerate mild calibration errors.
HF Daily Papers' note
G³VLA adds camera-aware geometry and cross-view fusion to pretrained vision-language-action models without changing their action space or training objective. On π₀, it improved LIBERO and RoboCasa24 performance using either ground-truth geometry or teacher predictions; the latter requires no depth sensors or manual geometric annotations. Camera-shifted simulations and real-robot experiments showed improved viewpoint robustness. Calibration tests found that the gains depend on correct camera geometry but tolerate mild calibration errors.
HF Daily Papers' note
score 5