FlowDrive:
moderated flow matching with data balancing for trajectory planning

News

• 2026/02/17 Initial code release
• 2025/09/26 FlowDrive paper published on arXiv.

Demo Figures & Videos

All demonstration GIFs correspond to figures in the paper.

Figure 7: Challenging Scenarios on nuPlan (Rows 1–4)

Row 1	Row 2
Row 3	Row 4

Figure 8: Challenging Scenarios on InterPlan (Rows 1–2)

Row 1	Row 2

Getting Started

• Setup the nuPlan dataset following the offiical-doc

Install dependencies

Install nuplan-devkit using the following commands or see the official doc

git clone https://github.com/motional/nuplan-devkit.git && cd nuplan-devkit
conda env create -f environment.yml
conda activate nuplan 
python -m pip install -e .

• Depending on your nuplan dataset location, you may need to add splits in the following two config files accordingly:
  • In nuplan-devkit/nuplan/planning/script/config/common/scenario_builder/nuplan_challenge.yaml, change the data_root: ${oc.env:NUPLAN_DATA_ROOT}/nuplan-v1.1/test/ to data_root: ${oc.env:NUPLAN_DATA_ROOT}/nuplan-v1.1/splits/test/.
  • In nuplan-devkit/nuplan/planning/script/config/common/scenario_builder/nuplan.yaml, change data_root: ${oc.env:NUPLAN_DATA_ROOT}/nuplan-v1.1/trainval to data_root: ${oc.env:NUPLAN_DATA_ROOT}/nuplan-v1.1/splits/trainval.

setup InterPlan devkit

cd ..
git clone https://github.com/mh0797/interPlan.git && cd interPlan
python -m pip install -e .

• You may need to change the config file interPlan/interplan/planning/script/config/common/scenario_builder/interplan.yaml, change data_root: ${oc.env:NUPLAN_DATA_ROOT}/nuplan-v1.1/trainval to data_root: ${oc.env:NUPLAN_DATA_ROOT}/nuplan-v1.1/splits/test.

setup TuPlan Garage

cd ..
git clone https://github.com/autonomousvision/tuplan_garage.git && cd tuplan_garage
python -m pip install -e .

• Replace the code tuplan_garage/tuplan_garage/planning/simulation/planner/pdm_planner/observation/pdm_object_manager.py by our code in flow_drive_planner/assets/adapted_tuplan_code/pdm_object_manager.py. The main fix with this code is that the objects predicted with constant velocity are now moving along their velocity direction, instead of the heading direction. This fix improves the performance of PDM and FlowDrive* significantly.

setup FlowDrive Planner

cd flow_drive_planner && python -m pip install -r requirements.txt
python -m pip install -e .

Closed-loop Evaluation

• Modify the evaluation script sim_flow_drive_planner.sh accordingly
• Run the evaluation script

chmod +x sim_flow_drive_planner.sh
./sim_flow_drive_planner.sh

Visualize the evaluation results

• Modify the visualization script run_nuboard_flow.py accordingly
• Run the visualization script python run_nuboard_flow.py

Training

• Configure the dara processing script data_process.sh and preprocess the training data

chmod +x data_process.sh
./data_process.sh

• Configure the paths in flow_drive_planner/flow_drive/config/config.yaml accordingly
• Run the training code. Note that the first time you train with weighted_sampling=True in flow_drive_planner/flow_drive/config/config.yaml, it will take a while to compute the sampling weights and save the weights to the same folder as the processed training data.

chmod +x train.sh
./train.sh 0,1,2,3,4,5,6,7  # specify the GPU ids

License

All assets and code are under the Apache 2.0 license unless specified otherwise.

Citation

If you find this work useful, please consider citing:

@article{wang2025flowdrivemoderatedflowmatching,
      title={FlowDrive: moderated flow matching with data balancing for trajectory planning}, 
      author={Lingguang Wang and Ömer Şahin Taş and Marlon Steiner and Christoph Stiller},
      year={2025},
      eprint={2509.21961},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2509.21961}, 
}

Acknowledgement

FlowDrive Planner is greatly inspired by the following open-source projects and we reuse some code from them : Diffusion-Planner, nuplan-devkit, tuplan_garage, planTF, pluto, DiT