projects

This project implements a Denoising Diffusion Probabilistic Model (DDPM) to learn a prior distribution over smooth, collision-free robot trajectories from expert demonstrations. The system consists of three main components: expert trajectory generation, model training, and inference.

1. Expert Trajectory Generation

We create a dataset of expert trajectories using A* search algorithm on a grid-based map with obstacles. These trajectories serve as the ground truth for training the diffusion model.

Key features:

• Creates a binary obstacle map with walls and barriers
• Generates start and goal points in free spaces
• Uses A* search to find optimal paths
• Normalizes and processes trajectories to a consistent format
• Saves the dataset with trajectories and their conditions (start, goal, obstacle map)

2. Model Training

We implement a diffusion model training pipeline.

Key components:

• We load the expert trajectories and their conditions
• Implements a conditional UNet architecture that takes trajectory state, time embedding, and environmental conditions
• Trains the model to denoise random Gaussian noise into expert trajectories
• Uses classifier-free guidance (CFG) for improved conditioning
• Saves model checkpoints during training

3. Inference

Key features:

• Loads trained model checkpoints
• Implements the reverse diffusion process with classifier-free guidance
• Generates smooth, collision-free trajectories between any valid start and goal
• Visualizes the results with the obstacle map

Denoising Process Visualization

The diffusion model works by gradually denoising a random Gaussian noise distribution into a coherent trajectory. The following visualization shows this process: