Diffusion-Based Planning

Diffusion-based planning studies how to generate useful action sequences and trajectories by refining noisy candidates, often conditioned on goals, maps, observations, or past actions. It is used in autonomous driving, robot navigation and manipulation, multi-robot coordination, and offline reinforcement learning, where multiple plausible plans can help handle uncertainty and complex constraints. Common approaches add guidance from rewards, learned dynamics, safety costs, or language and world models; hierarchical planning, trajectory composition, and distillation aim to extend planning horizons and reduce inference time. A central challenge is making diverse generated plans reliably safe, feasible, and fast in closed-loop settings, while preserving the ability to generalize beyond training data.