---

name: uniad description: "Operate the UniAD repository for planning-oriented autonomous driving data preparation, configs, training/evaluation, checkpoints, and visualization workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

UniAD repo skill

Use this skill when a task involves the OpenDriveLab UniAD repository, planning-oriented autonomous driving, nuScenes-based BEV perception/prediction/planning workflows, UniAD checkpoints, OpenMMLab configs, projects.mmdet3d_plugin, or UniAD train/eval/visualization scripts.

UniAD is not a normal installed Python distribution in this checkout. It is an OpenMMLab plugin rooted at projects.mmdet3d_plugin, with public workflows driven by config files and launcher scripts.

Quick route map

User intent	Read
Prepare or validate nuScenes, CAN bus, map extensions, info PKLs, or motion anchors	sub-skills/data-preparation/SKILL.md
Choose/edit a config, understand stage1 vs stage2, inspect model heads/classes, or debug plugin/model registry issues	sub-skills/config-and-model-architecture/SKILL.md
Build train/eval/SLURM commands, place checkpoints, plan GPU usage, resume runs, or reproduce metrics	sub-skills/training-evaluation/SKILL.md
Inspect results.pkl, render BEV/camera/video outputs, or debug visualization/log artifacts	sub-skills/visualization-and-results/SKILL.md

Start here for new environments

Read references/installation.md before giving install advice. The public v2.0 docs specify Python 3.9, Torch 2.0.1+cu118, mmcv-full==1.6.1, mmdet==2.26.0, mmsegmentation==0.29.1, mmdet3d==1.0.0rc6, then requirements.txt.

Minimal plugin check from a UniAD checkout:

PYTHONPATH="$(pwd)":$PYTHONPATH python - <<'PY'
import torch, mmcv, mmdet, mmseg, mmdet3d
print(torch.__version__, torch.version.cuda, torch.cuda.is_available())
print(mmcv.__version__, mmdet.__version__, mmseg.__version__, mmdet3d.__version__)
import projects.mmdet3d_plugin
print('UniAD plugin import ok')
PY

Use scripts/check_uniad_environment.py when you need a structured import/CUDA/config smoke check.

Core workflow order

1. Verify the environment and plugin import.
2. Use data-preparation to confirm data/nuscenes/, data/infos/, and stage-2 motion anchors.
3. Use config-and-model-architecture to choose between BEVFormer, stage1 track/map, and stage2 E2E configs.
4. Use training-evaluation to render train/eval commands and check checkpoints/GPU expectations.
5. Use visualization-and-results only after evaluation has produced a result pickle or show directory.

Repo-level references

• references/repo-provenance.md — source commit, branch, evidence paths, and refresh triggers.
• references/repo-routing-metadata.json — managed repo-skills-router metadata.
• references/installation.md — public runtime stack, plugin import model, and version caveats.
• references/checkpoints-and-models.md — stage flow, checkpoint names, and task ownership.
• references/troubleshooting.md — cross-cutting symptom-to-route triage.

Bundled repo-level script

• scripts/check_uniad_environment.py — checks imports, CUDA availability, plugin import, and optionally parses the public configs. Run it from a UniAD checkout or pass --repo-root.

Important constraints

• Do not claim full metric reproduction from import/config checks alone. Full reproduction requires the nuScenes data, CAN bus/map extensions, correct info PKLs, checkpoints, compatible CUDA/OpenMMLab packages, and enough GPUs/VRAM.
• Do not tell future agents to run original repo scripts as the only guidance. Use the bundled command builders and references in this skill to synthesize safe commands.
• The legacy Dockerfile targets older CUDA/Torch/OpenMMLab versions than the v2.0 docs; treat it as historical unless the user asks for legacy setup.
• UniAD training/evaluation is expensive. For exploratory triage, prefer layout validation, config parsing, command rendering, and tiny import/CUDA checks before launching jobs.