JAMB: Joint Action-Motion Diffusion for Bimanual Manipulation

JAMB addresses coordinated bimanual manipulation where each arm's motion affects the shared 3D scene. Unlike prior diffusion policies that generate actions without explicit future geometric consequences, JAMB jointly denoises both actions and motion states. The approach enables robots to anticipate how their movements will impact the scene and affect the other arm's actions, improving coordination in dual-arm manipulation tasks.