pengzhenghao97/agenticnav-mini

AgenticNav — mini split

Mini fixture for AgenticNav, the dataset accompanying the paper "Slow Brain, Fast Planner: Latency-Resilient VLM-Augmented Urban Navigation".

This split contains 1 episode (s2e_v2_20260115_224153) with 3 takeover clips for fast CI / quickstart smoke tests. For full evaluation, use the agenticnav-hard split.

Quickstart

git clone https://github.com/pengzhenghao/AgenticNav
cd AgenticNav
uv sync && source .venv/bin/activate
python scripts/download_dataset.py --split mini   # writes data/agenticnav-mini/
python -m agentnav.cli.trajectory_selection \
    --dataset data/agenticnav-mini --model dummy_argmax --write-report

Layout (canonical episode schema v0.1.0)

agenticnav-mini/
├── dataset_manifest.json
├── episodes/
│   └── s2e_v2_20260115_224153/
│       ├── episode.json
│       ├── rgb.jsonl                  # video-backed RGB stream (5 Hz)
│       ├── odom.jsonl                 # robot pose in map ENU frame
│       ├── planner_candidates.jsonl   # S2E candidate trajectories + scores per tick
│       └── assets/rgb/
│           ├── front.mp4              # canonical RGB video
│           ├── front_pinhole.mp4      # pinhole-projected RGB (for planner input)
│           └── sample_*.png           # individual frames referenced by takeover_clips
└── takeover_clips/
    ├── takeover_clips.jsonl           # 3 clips
    └── summary.json

See the code repo for the full schema (src/agentnav/schema/canonical_episode.py).

License

MIT. See LICENSE in the code repo.

Citation

@inproceedings{peng2026slowbrain,
  title={Slow Brain, Fast Planner: Latency-Resilient VLM-Augmented Urban Navigation},
  author={Peng, Zhenghao and others},
  booktitle={TODO},
  year={2026}
}