MOSAIC: Mode- and Support-Aware Inference-Time Control of Robot Diffusion Models

MOSA, a mode- and support-aware framework for gradient-free inference-time control of robotic diffusion models, treats initial noise and stochastic reverse-step perturbations as a joint input to the pretrained denoising process. It adapts a distribution over this input using importance weights computed from fully denoised outputs, balancing physical feasibility and consistency with the learned support through adaptive penalties. MOSAIC enables coordinated behavior across multiple independently trained diffusion models without requiring a joint model, and is evaluated on navigation, dexterous grasping, motion planning with unseen geometry, and collaborative box manipulation.