Developed a learning-based framework for 2D path planning in continuous environments with polygonal obstacles using Denoising Diffusion Probabilistic Models (DDPMs). The diffusion model learns to generate smooth, goal-directed, collision-free trajectories without requiring explicit search at inference time.
Traditional planners like A* or RRT generate one deterministic path and often require heavy obstacle checking. Diffusion models can learn the distribution of feasible paths and sample diverse, smooth trajectories, enabling faster and more flexible planning for robotics and autonomous navigation.
Environment: 2D continuous space with Shapely-modeled convex and non-convex obstacles.
Data Generation: Collision-free paths generated via a basic RRT planner from random start/goal pairs.
Trajectory Representation: Flattened, normalized 2D point sequences.
Model: Lightweight MLP-based DDPM conditioned on start and goal positions.
Training: Thousands of noisy trajectories, simple linear noise schedule.
Sampling: Iterative denoising to produce realistic paths.
Smoothness: Generated trajectories are smoother than raw RRT paths.
Goal-Reaching: Conditioned paths consistently reach goals.
Obstacle Avoidance: Paths avoid obstacles implicitly.
Flexibility: Can handle diverse start/goal queries without retraining.