CBF Diffusion

Learning and validating Control Barrier Functions (CBFs) on a double-integrator system.

Project Layout

.
├── artifacts/
│   ├── data/          # Generated trajectory files (.pt)
│   └── models/        # Trained models (.pth/.onnx)
├── legacy/
│   ├── experiments/   # Older exploratory scripts
│   ├── matlab/        # Original MATLAB scripts
│   └── data/          # Legacy .mat files
├── diffusion_cbf.py   # Diffusion-based CBF training/evaluation
├── trainingDataGen.py # Generate safe trajectories (Python replacement)
├── trainNN.py         # Train neural CBF regressor
├── validateNN.py      # Validate learned CBF in closed-loop simulation
└── Makefile

Setup

python3 -m pip install -r requirements.txt

Standard CBF Workflow

python3 trainingDataGen.py
python3 trainNN.py
python3 validateNN.py

Equivalent make targets:

make generate-data
make train-cbf
make validate-cbf

Diffusion-Based CBF Workflow

python3 diffusion_cbf.py --epochs 300 --device cpu

This prints:

• diffusion training loss,
• safe/unsafe classification metrics (accuracy, precision, recall, f1),
• closed-loop safety metrics (unsafe_state_ratio, projection_infeasible_ratio).