This project builds upon the "Planning with Diffusion for Flexible Behavior Synthesis" framework, advancing its capabilities for trajectory planning in goal-conditioned reinforcement learning (RL) and addressing limitations in traditional model-based approaches. Diffusion models, known for their iterative denoising process, inherently provide scalability and temporal compositionality, making them an ideal choice for synthesizing flexible and robust plans.
We propose key extensions to the Diffuser framework, including middle waypoint conditioning, multi-agent trajectory planning, and dynamic obstacle avoidance. By conditioning on additional constraints such as intermediate waypoints and dynamic obstacle positions, the framework generates globally coherent, context-aware trajectories that adapt to complex environments. Furthermore, the multi-agent planning extension ensures safe and coordinated movement in collaborative settings, a critical need in robotics and autonomous systems.
Our experiments in the Maze2D task demonstrate significant improvements in trajectory feasibility, safety, and planning efficiency. The results showcase the model's ability to navigate challenging scenarios with reduced suboptimality, increased obstacle clearance, and enhanced multi-agent coordination. These advancements highlight the potential of diffusion-based planning for applications requiring robust and flexible trajectory synthesis in dynamic, multi-agent environments.