SuReNav

Superpixel graph-based constraint relaxation for navigation in over-constrained environments.

data_gathering/   interactive demonstration collection
data_processing/  superpixel segmentation, postprocessing, scenario perturbation
model/            GraphGPS training package
eval/             evaluation drivers and metrics
_common/          shared labels and helpers

Per-stage details live in each subdirectory's README.md.

---

a. Installation

git clone <REPO_URL> surenav && cd surenav

conda create -n constrelax python=3.10 && conda activate constrelax

# PyTorch + PyG (CUDA 11.7)
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 \
    --extra-index-url https://download.pytorch.org/whl/cu117
pip install pyg-lib torch-scatter torch-sparse torch-cluster torch-spline-conv \
    -f https://data.pyg.org/whl/torch-1.13.1+cu117.html
pip install torch_geometric==2.2.0

pip install -r requirements.txt

All commands assume the repo root as the working directory.

b. Generating semantic maps

We use MapItAnywhere to produce per-pixel semantic masks. Align the class taxonomy with _common/labels.py and export one <MAP_NAME>.npy per location.

c. Dataset gathering

Preprocess the semantic masks, then collect demonstrations:

python -m data_processing.superpixel_segmentation ...
python -m data_processing.merge_superpixels ...

python -m data_gathering.gather_traj_data \
    --base_path      <SEM_DIR> \
    --region_map_dir <SP_DIR> \
    --out            <OUT_DIR> \
    --samples 10

d. Training

cd model
CUDA_VISIBLE_DEVICES=0 python main.py --cfg configs/constrelax.yaml

See model/README.md for the dataset layout.

e. Evaluation

Inference-only against a pretrained checkpoint (--checkpoint_dir).

Quantitative:

python -m eval.run_quantitative --checkpoint_dir <CKPT_DIR> ...
python -m eval.metrics.quantitative --log_dirs <LOG_DIR> ...

Human-likeness:

python -m eval.run_humanlike --checkpoint_dir <CKPT_DIR> ...
python -m eval.metrics.humanlikeness --log_dirs <LOG_DIR> ...

See eval/README.md for the full CLI.

---

Acknowledgements

Built on the shoulders of:

• GraphGPS — graph transformer backbone.
• Neural A* — differentiable A* head.
• MapItAnywhere — semantic map generation.

Citation

@inproceedings{koh2026surenav,
  title     = {SuReNav: Superpixel Graph-based Constraint Relaxation for Navigation in Over-constrained Environments},
  author    = {Koh, Keonyoung and Jung, Moonkyeong and Lee, Samuel Seungsup and Park, Daehyung},
  booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
}