Dexterity-BEV: Aligning 3D World and Actions for Generalizable Robot Policy Learning

Researchers introduced Dexterity-BEV, a robot manipulation framework that aligns camera views and actions in a shared 3D coordinate system.

The method adds pixel-wise 3D inputs, bird's-eye-view image construction, and temporal trajectory alignment across robots and datasets.

By putting perception and action into one common frame, the approach aims to improve policy generalization across embodiments, viewpoints, and tasks.

Code and data tools were released for training and evaluation.