Can robots learn dexterous skills by aligning 3D space across cameras and bodies?

Robots trained on end-to-end manipulation often fail when camera angles or robot bodies change because policies learn from 2D images without spatial grounding. This work adds 3D awareness by computing pixel-wise 3D coordinates from camera calibration, then aligning both visual inputs and robot actions to a shared bird's-eye-view frame. A temporal alignment scheme also handles different recording speeds across robots and datasets. The method improves consistency and real-world transfer; code, trained models, and data pipeline are released.