PC-Diffuser: Path-Consistent Capsule CBF Safety Filtering for Diffusion-Based Trajectory Planner

Eugene Ku, Yiwei Lyu

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026

PC-Diffuser is a context-aware safeguarding mechanism that evaluates collision risk using a capsule-distance barrier function that better reflects vehicle geometry. Based on this collision risk score, a path-consistent safety filter is applied which prevents geometric distortion from the nominal trajectory, minmizing invasive distribution shift from the learned model. PC-diffuser injects these safety-consistent corrections at every denoising step, enabling iterative, context-aware safeguarding instead of post-hoc repair.

Getting Started

1. Download data

Sign up at https://www.nuscenes.org/nuplan and download nuPlan Val Split.

Then set the data path appropriately in run_all.sh and run_vis_examples.sh.

You may need to change the name of the data dir to trainval.

2. Set-up repo and environment

git clone --recursive https://github.com/Eugene29/PC-Diffuser.git

conda create -n pc-diffuser python=3.9 -y
conda activate pc-diffuser

# Install appropriate torch for your GPU
pip install torch==2.0.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install torchvision==0.15.1+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

pip install "pip<24.1"
pip install --no-build-isolation grpcio==1.43.0
pip install -r requirements.txt

pip install -e nuplan-devkit
pip install -e diffusion-planner-cbf 

# download weights
mkdir -p diffusion-planner-cbf/checkpoints
wget -P diffusion-planner-cbf/checkpoints https://huggingface.co/ZhengYinan2001/Diffusion-Planner/resolve/main/args.json
wget -P diffusion-planner-cbf/checkpoints https://huggingface.co/ZhengYinan2001/Diffusion-Planner/resolve/main/model.pth

3. Run reproducers

To reproduce the table 1 of our paper

bash scripts/run_all.sh

To reproduce the figure 1 and supplementary visualizations

bash scripts/run_vis_examples.sh

Figure 1: Check out the video demo for other two comparison scenarios.

4. Key folders and files

1. Bulk of the safety related code lives in Diffusion_Planner_CBF/diffusion_planner/safety and Diffusion_Planner_CBF/diffusion_planner/model/diffusion_utils/dpm_solver_pytorch.py
2. For mpc-cbf and pc-diffuser, we implemented a custom action replay tracker so that CBF's action output is not overriden by nuplan's controller. Check out nuplan-devkit/nuplan/planning/simulation/controller/tracker/action_replay_tracker.py

Bibtex

If you find our paper and code useful, please cite our paper as:

@inproceedings{ku2026pc,
  title={PC-Diffuser: Path-Consistent Capsule CBF Safety Filtering for Diffusion-Based Trajectory Planner},
  author={Ku, Eugene and Lyu, Yiwei},
  booktitle={2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2026}
}

Acknowledgement

PC-Diffuser is greatly inspired by the following outstanding contributions to the open-source community: nuplan-devkit, Diffusion-Planner, tuplan_garage, pluto, DiT, dpm-solver.