Learning Geometrically-Grounded 3D Visual Representations for View-Generalizable Robotic Manipulation

Real-world robotic manipulation demands visuomotor policies capable of robust spatial scene understanding and strong generalization across diverse camera viewpoints. While recent advances in 3D-aware visual representations have shown promise, they still suffer from several key limitations: (i) reliance on multi-view observations during inference, which is impractical in single-view restricted scenarios; (ii) incomplete scene modeling that fails to capture holistic and fine-grained geometric structures essential for precise manipulation; and (iii) lack of effective policy training strategies to retain and exploit the acquired 3D knowledge. To address these challenges, we present GEM3D (Geometrically-Grounded 3D Manipulation), a unified representation-policy learning framework for view-generalizable robotic manipulation. GEM3D introduces a single-view 3D pretraining paradigm that leverages point cloud reconstruction and feed-forward gaussian splatting under multi-view supervision to learn holistic geometric representations. During policy learning, GEM3D performs multi-step distillation to preserve the pretrained geometric understanding and effectively transfer it to manipulation skills. We conduct experiments on 12 RLBench tasks, where our approach outperforms the previous state-of-the-art (SOTA) method by 12.7% in average success rate. Further evaluation on six representative tasks demonstrates the strong zero-shot view generalization of our approach, with the success rate drops by only 22.0% and 29.7% under moderate and large viewpoint shifts, respectively, whereas the SOTA method suffers larger decreases of 41.6% and 51.5%.