You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These scene reconstructions are derived from nuPlan sensor data. They are shared for non-commercial research only, under CC BY-NC-SA 4.0 and the nuPlan terms of use (https://www.nuscenes.org/terms-of-use).

Log in or Sign Up to review the conditions and access this dataset content.

Odyssey Scenes

Scene data for the Odyssey closed-loop driving benchmark: every file the Odyssey simulator reads for each scene.

  • Each scene is a 100-second segment of a nuPlan log.
  • The camera images are rendered from an OmniRe reconstruction (3D Gaussian splatting with dynamic actors). The rest of the scene is also included:
    • the log's trajectories
    • the road surface
    • the route
    • traffic-light labels
  • These are simulator inputs, not standalone models. The simulator code and the inputs shared by all scenes (nuPlan maps, the image restorer) are distributed separately.

Contents

There is one folder per scene, odyssey_scene001/ … odyssey_scene100/:

File What Size, 100 scenes
scene.ckpt OmniRe reconstruction: the static background and every actor 209.9 GB (1.1–5.2 GB each)
scenario.pkl the scenario: ego and agent tracks, traffic-light states and timestamps from the log 7.1 GB
road_surface.npz the road surface the renderer places actors on 53 MB
road_surface.json road-surface parameters, required: the renderer reads drop_static from it (an actor that moves less than this many metres in the scene stays where the reconstruction put it instead of being set onto the road surface). Without it the rendering changes. 75 KB
route.npz the route: input of route-conditioned planners and reference for route-completion scoring 4.3 MB
tl_control.json traffic-light labels: signal heads and the lane connectors they control 0.7 MB
checkpoint_inventory.json table of contents of scene.ckpt: node types and tensor shapes. The benchmark's run configuration reads it to decide whether a scene gets traffic-light labels. 1.3 MB

Files at the top level:

  • scenes.csv: one row per scene, with these columns:
    • scene: the folder in this repo.
    • simulator_asset_id, token, scenario_dir: the names the simulator uses for the scene.
    • city, nuplan_log: where the scene comes from.
    • start_timestamp_us, end_timestamp_us: the segment's start and end, in microseconds.
    • frames: frame count.
  • MD5SUMS: checksums of every file, for md5sum -c.

Coverage: 100 segments from 63 nuPlan logs. Each segment is 1,000 frames at 10 Hz (~100 s).

City Scenes
Las Vegas (us-nv-las-vegas-strip) 32
Boston (us-ma-boston) 29
Singapore (sg-one-north) 21
Pittsburgh (us-pa-pittsburgh-hazelwood) 18

Download

Access is gated. Log in to Hugging Face and accept the terms on this page first.

hf auth login

# all 100 scenes
hf download ADRLAB/odyssey-scenes --repo-type dataset --local-dir odyssey-scenes

# a single scene (plus the index and checksums)
hf download ADRLAB/odyssey-scenes --repo-type dataset --local-dir odyssey-scenes \
  --include "odyssey_scene001/*" --include scenes.csv --include MD5SUMS
  • An interrupted download resumes when you run the same command again.

Verify:

cd odyssey-scenes
md5sum -c MD5SUMS                    # all scenes
md5sum -c --ignore-missing MD5SUMS   # after a partial download

Install into the simulator

The simulator finds each file under its own names, listed in scenes.csv. Link the downloaded files into that layout:

cd odyssey-scenes
D=<ODYSSEY_ROOT>/data/sim_engine   # <ODYSSEY_ROOT>: your simulator checkout
tail -n +2 scenes.csv | while IFS=, read -r scene asset token sdir _; do
  [ -d "$scene" ] || continue      # skip scenes you did not download
  S="$PWD/$scene"; c="${asset%%_*}"
  mkdir -p "$D/assets/omnire_v11_140k/$asset/background" "$D/scenarios/omnire_v11_140k_actorpack_v1/$sdir" \
           "$D/anchor_surface_nostatic_v11" "$D/route_sidecars_alt/label_v11" "$D/tl_control_v11/$c" \
           "$D/assets/omnire_v11_140k/$asset/provenance"
  ln -sf "$S/scene.ckpt"        "$D/assets/omnire_v11_140k/$asset/background/$asset.ckpt"
  ln -sf "$S/scenario.pkl"      "$D/scenarios/omnire_v11_140k_actorpack_v1/$sdir/all_scenarios.pkl"
  ln -sf "$S/road_surface.npz"  "$D/anchor_surface_nostatic_v11/$token.npz"
  ln -sf "$S/road_surface.json" "$D/anchor_surface_nostatic_v11/$token.json"
  ln -sf "$S/route.npz"         "$D/route_sidecars_alt/label_v11/$token.npz"
  ln -sf "$S/tl_control.json"   "$D/tl_control_v11/$c/tl_control.json"
  ln -sf "$S/checkpoint_inventory.json" "$D/assets/omnire_v11_140k/$asset/provenance/checkpoint_inventory.json"
done

Files that already exist at those paths are replaced by links.

License

Citation

If you use these scenes, please cite nuPlan and OmniRe:

@article{caesar2021nuplan,
  title   = {nuPlan: A closed-loop ML-based planning benchmark for autonomous vehicles},
  author  = {Caesar, Holger and others},
  journal = {arXiv preprint arXiv:2106.11810},
  year    = {2021}
}

@inproceedings{chen2025omnire,
  title     = {OmniRe: Omni Urban Scene Reconstruction},
  author    = {Chen, Ziyu and others},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2025}
}
Downloads last month
38

Papers for ADRLAB/odyssey-scenes