Skip to content
zerlinwangPublic

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

Learning to Drive in New Cities Without Human Demonstrations

NeurIPS 2026

Project Page arXiv Dataset

Zero-shot transfer vs. NOMAD under Singapore-to-Boston transfer

This is the official implementation of NOMAD (NO data Map-based self-play for Autonomous Driving), which adapts a driving policy to a new city using only the target-city map and meta-information, without any human demonstrations from that city. Above: under Singapore-to-Boston transfer, (a) the zero-shot policy keeps driving on the left, causing collisions and off-road events, while (b) NOMAD keeps to the right and all vehicles reach their goals.

Environment Configuration

Compile Docker image

DOCKER_BUILDKIT=1 docker build --build-arg USE_CUDA=true --tag gpudrive:latest --progress=plain .

Install dependencies

# Run Docker container
docker run --gpus all -it --rm --shm-size=20G -v ${PWD}:/workspace gpudrive:latest /bin/bash
# Build gpudrive
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_POLICY_VERSION_MINIMUM=3.5 && find external -type f -name "*.tar" -delete
make -j

Dataset Download

NuPlan scenarios (Pittsburgh, Singapore, Boston) are available as the NOMAD dataset on Hugging Face.

Data Preparation

The training pipeline consumes GPUDrive-format JSON scenarios. Two preparation stages turn raw nuPlan logs into the clean subset used in our experiments:

1. Convert nuPlan → GPUDrive JSON (conversion/)

export NUPLAN_DATA_ROOT=/path/to/nuplan/dataset   # directory with .db files
export NUPLAN_MAPS_ROOT=/path/to/nuplan/maps

cd conversion
make nuplan
scenariomax-convert \
    --nuplan_src $NUPLAN_DATA_ROOT \
    --dst /path/to/output \
    --target_format gpudrive \
    --num_workers 16 \
    --nuplan_scenario_duration 15.0

See conversion/README.md and conversion/docs/DATASETS.md for the full schema and CLI options.

2. Filter to valid scenarios (filtering/)

A scenario is kept only if it has no off-road incidents, runs for more than 5 simulation steps, and contains at least one controllable agent.

# Run the combined filter -- emits valid_scenarios_only_<folder>.json under ./output/
python -m filtering.scenarios_valid_filter <data_path> <max_files>

# Copy the valid scenario files into a clean directory
python -m filtering.copy_filtered_scenarios

Training

Demonstration Generation

python baselines/bc/collect_multi_discrete_demo.py

Imitation Learning

python baselines/bc/bc.py

Scenario Generation

python baselines/goal/generate_scene_heuristic_modular.py

Reinforcement Learning

python baselines/bcsp/ppo_finetuning.py

WOSAC Evaluation

python baselines/eval/wosac_evaluate.py

Project Page

The source of the project page lives in docs/ and is served with GitHub Pages.


If you find this repo helpful, don't forget to give us a star! If this repo helps your research, please cite this work:

@inproceedings{wang2026learning,
  title={Learning to Drive in New Cities Without Human Demonstrations},
  author={Wang, Zilin and Rahmani, Saeed and Cornelisse, Daphne and Sarkar, Bidipta and Goldie, Alexander David and Foerster, Jakob Nicolaus and Whiteson, Shimon},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}

@inproceedings{kazemkhani2025gpudrive,
      title={GPUDrive: Data-driven, multi-agent driving simulation at 1 million FPS},
      author={Saman Kazemkhani and Aarav Pandya and Daphne Cornelisse and Brennan Shacklett and Eugene Vinitsky},
      booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
      year={2025},
      url={https://arxiv.org/abs/2408.01584},
      eprint={2408.01584},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
}

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages