Can robots learn to move by reusing motion building blocks?

Robot trajectories contain recurring motion fragments—reach, grasp, place—that could be reused like Lego blocks. This work treats trajectory generation as assembly of learned motion primitives in physical space, not abstract latent codes. A geometric loss ensures smooth transitions between primitives. On Open X-Embodiment and 3DMoTraj benchmarks, the method improves final displacement error by 21% and average displacement by 19%, while cutting prediction drift ratio from 1.8 to 1.07.