Model-Based Diffusion Sampling for Predictive Control in Offline Decision Making
Researchers have developed a new framework called Model Predictive Diffuser (MPDiffuser) to improve the reliability of diffusion models in offline decision-making tasks. This approach combines a diffusion planner with a dynamics diffusion model, allowing for the generation of trajectories that are both aligned with task objectives and dynamically plausible. MPDiffuser iteratively refines feasibility while maintaining task intent, and a ranking module selects the best trajectories. The framework has shown consistent improvements over previous diffusion-based methods on various benchmarks and has been validated on a real quadrupedal robot. AI
IMPACT Enhances the reliability of diffusion models for real-world control tasks, potentially improving robotic applications.