lvchangfu/streetworld-benchmark

StreetWorld Benchmark

StreetWorld provides reconstructed driving scenes for closed-loop evaluation of autonomous driving policies. This release contains 650 nuScenes scenes represented with dynamic 3D Gaussian splatting, together with their scene metadata and a shared HD map cache.

The scenes were reconstructed using Hierarchy UGP and packaged for the default renderer in StreetWorld.

Release contents

Component	Contents
Scene archives	650 .npz files from nuScenes, one per scene
Maps	One shared nuscenes/map_cache.npz
Total asset size	424,176,255,753 bytes (395.04 GiB)
Archive format	StreetWorld NPZ format version 1
Source locations	Boston and Singapore

The 650 scenes form StreetWorld's nuScenes benchmark selection. Scene IDs retain their original four-digit names. This release does not assign new training, validation, or test splits.

streetworld-benchmark/
├── README.md
├── LICENSE
└── nuscenes/
    ├── scene_list.txt
    ├── map_cache.npz
    ├── 0001.npz
    ├── 0002.npz
    └── ...

Download and use

Download the nuScenes assets with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="lvchangfu/streetworld-benchmark",
    repo_type="dataset",
    allow_patterns=["nuscenes/*"],
    local_dir="/path/to/streetworld-benchmark",
)

After installing StreetWorld and its ST Renderer (default renderer), point the simulator to the directory containing the archives:

from st_renderer import SimulatorInterface

simulator = SimulatorInterface(
    "nuscenes",
    root="/path/to/streetworld-benchmark/nuscenes",
)

To download a single scene, retrieve both the scene archive and the shared map cache:

from huggingface_hub import hf_hub_download

for filename in ["nuscenes/0001.npz", "nuscenes/map_cache.npz"]:
    hf_hub_download(
        repo_id="lvchangfu/streetworld-benchmark",
        repo_type="dataset",
        filename=filename,
        local_dir="/path/to/streetworld-benchmark",
    )

Archive format

Each scene archive can be opened with numpy.load(path, allow_pickle=False) and contains the following members:

Key	Contents
format_version	Integer archive version, currently 1
scene	UTF-8 JSON bytes with the dataset, scene ID, timestamp range, and map ID
camera	UTF-8 JSON bytes with camera intrinsics, image dimensions, and ego-to-camera transforms
ego_pose	UTF-8 JSON bytes with timestamped ego poses
bounding_box	UTF-8 JSON bytes with object types, dimensions, and timestamped poses
object_offset	UTF-8 JSON bytes with object pose offsets
background	Byte array containing the reconstructed background PLY
objects	Byte array containing the object model PTH payload
skymap	Byte array containing the sky model PTH payload

The shared map archive contains format_version, a JSON manifest, and vectorized map payloads. It covers:

• nusc_trainval:boston-seaport
• nusc_trainval:singapore-hollandvillage
• nusc_trainval:singapore-onenorth
• nusc_trainval:singapore-queenstown

License and attribution

The nuScenes-derived assets, including the reconstructed scenes and map cache, are provided under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0), subject to the additional nuScenes Dataset Terms of Use. See LICENSE for the license text and attribution notice.

The source data and maps were created by nuTonomy / Motional. Please credit both StreetWorld and nuScenes, retain the license and attribution notices, and identify changes when redistributing adaptations. Commercial use requires permission under the upstream terms.

StreetWorld is an independent project; this release does not imply endorsement by Motional. The StreetWorld source code has its own Apache 2.0 license.