WJCUCL/sandplanner-dataset-avoid

SanD-Planner training data — dataset_avoid

The dataset_avoid subset (obstacle-avoidance runs) used to train the SanD-Planner local trajectory planner.

• Collected in a custom Gazebo simulation environment (self-built).
• This is one subset of a larger training set — only part of the full dataset is released here.

Contents

A single archive dataset_avoid.zip (~14 GB) that extracts to:

dataset_avoid/
└── run_XXXX/                # 152 runs, variable length
    ├── traj_xyz.npy         # (T, 3) trajectory positions (x, y, z), metres
    ├── traj_yaw.npy         # (T,)   heading (yaw)
    ├── traj_pitch.npy       # (T,)   pitch
    └── depth/
        └── depth_NNNN.png   # per-frame depth images (uint16, mm)

Usage

hf download WJCUCL/sandplanner-dataset-avoid dataset_avoid.zip --repo-type dataset --local-dir .
unzip dataset_avoid.zip -d dataset/        # -> dataset/dataset_avoid/run_XXXX/
DATASET_ROOT=dataset bash run_train.sh     # run inside the SanD-Planner repo

See the SanD-Planner repository for the training and inference code.

Citation

@article{wang2026sand,
  title   = {SanD-Planner: Sample-Efficient Diffusion Planner in B-Spline Space for Robust Local Navigation},
  author  = {Wang, Jincheng and Bao, Lingfan and Yang, Tong and Plasencia, Diego Martinez and Jiao, Jianhao and Kanoulas, Dimitrios},
  journal = {arXiv preprint arXiv:2602.00923},
  year    = {2026}
}