OccStress

Stress-Testing the 4D Occupancy Forecasting Chain
NeurIPS 2026 · Evaluations and Datasets Track

Download  ·  Coverage  ·  Data Format  ·  Licenses and Use

How do perception errors affect future occupancy forecasts? OccStress evaluates robustness across the occupancy forecasting chain, from sensor errors to 3D occupancy states and multi-horizon predictions.

3 datasets · 10,827 unique anchors · 975 protocols · 6 future horizons

This repository provides evaluation protocols, derived occupancy assets and supporting metadata. All 75 archives are available, totaling 40.95 GB compressed and 143.67 GB extracted including metadata. Download only the dataset, track or upstream source you need.

Original camera images, LiDAR, clean GT and model checkpoints are not bundled. See external prerequisites and source-dependent terms.

Dataset Coverage

Dataset	Source	Scenes	Anchors per protocol
OccStress-nuScenes	nuScenes / Occ3D	150	4,519
OccStress-Waymo	Waymo / Occ3D	202	5,978
OccStress-CARLA	UniOcc CARLA simulation	3	330

All datasets use 2 Hz observations, a canonical window of four historical states plus the current state, and six future targets at +0.5 to +3.0 seconds. The selected frames total 6,019 / 7,998 / 360 for nuScenes / Waymo / CARLA.

Published protocols	nuScenes	Waymo	CARLA
Manual	38	38	38
Upstream	364	219	218
Position sweep	60	0	0

Counts include clean references. Protocols reuse anchors and assets; 975 protocols do not represent 975 independent datasets or new sets of samples. Waymo's two EFFOcc sensor-stress branches retain separate clean protocol files.

What Can You Evaluate?

• Upstream robustness: use source-specific occupancy predictions under camera or LiDAR stressors to measure error propagation into forecasting.
• Manual robustness: test semantic replacement, dropout, holes, misalignment and traffic mirroring directly in occupancy states.
• Temporal sensitivity: compare Current-only, Recent-burst and History-only failure regimes, or use the fixed-budget, single-state position sweep to isolate input-position effects.

Traffic mirroring transforms inputs, targets and motion metadata consistently and is reported separately from non-traffic robustness. Four-state models do not observe the earliest canonical input; that position is marked unobserved, not robust. See format and evaluation conventions.

Download

Start With One Dataset and Track

For example, download the Waymo manual track and its metadata:

python -m pip install -U huggingface_hub
hf download insailab/OccStress --repo-type dataset \
  --local-dir /datasets/OccStress-download \
  --include README.md DOWNLOAD.md FORMAT.md EXTERNAL_DATA.md DISTRIBUTION.md packages.json \
  'metadata/common.tar.zst' 'metadata/OccStress-Waymo.tar.zst' \
  'archives/manual/OccStress-Waymo/*.tar.zst'

Use DOWNLOAD.md for other tracks, source-specific selection, pinned revisions, checksum verification and extraction commands. The package index lists every archive's size and SHA256.

Extract selected archives into one shared OccStress/ root, then prepare external GT and metadata before running a model. Reserve space for both downloaded archives and their extracted contents; original dependencies need additional storage.

What is inside the extracted dataset?

OccStress/
  meta/<dataset>/                   Dataset metadata and controls
  protocols/manual/<dataset>/      Manual evaluation protocols
  protocols/upstream/<dataset>/    Source-specific evaluation protocols
  protocols/position_sweep/        Single-state temporal diagnostics
  occ/manual/<dataset>/            Shared manual corruption assets
  occ/upstream/<dataset>/          Exported upstream occupancy
  events/manual/<dataset>/         Corruption event metadata

<dataset> is OccStress-nuScenes, OccStress-Waymo or OccStress-CARLA. Protocol paths are relative to this root. Position sweeps reuse the existing occupancy assets; they do not need duplicate voxel packs. See FORMAT.md.

Before Evaluation

• Use the matching ground truth. Native Waymo labels require the supplied mapping; CARLA requires the canonical right-handed view. Already converted occupancy arrays must not be transformed twice.
• Keep the model's input and control policy. Use all six genuine future targets; do not substitute a current-frame reconstruction.
• Keep metric conventions explicit. Historical paper metrics and the released present-class convention are not interchangeable. Paper averages use the 1, 2 and 3 second horizons.
• Use the assets directly when applicable. Occupancy-input forecasting does not require upstream weights or original sensors. Camera-direct methods and upstream re-export do require the relevant sensor inputs.

This is an archive-based benchmark, not an Arrow table for datasets.load_dataset. Only unpickle protocol files from trusted sources.

Scope and Limitations

OccStress is a controlled diagnostic stress test, not a deployment-risk estimate or safety certification. Its corruptions are reproducible stress conditions, not a calibrated distribution of real-world failures.

CARLA contains only three simulated validation scenes. The datasets differ in class support and sensor configuration; report cross-dataset and zero-shot settings explicitly. Published protocols do not imply that every model supports every track or that every model/protocol combination has been evaluated.

Preparation checks and their scope are recorded in FORMAT.md and the published manifests. Source-dataset registration, attribution and use conditions still apply; see DISTRIBUTION.md.

Citation

@inproceedings{zheng2026occstress,
  title = {{OccStress}: Stress-Testing the {4D} Occupancy Forecasting Chain},
  author = {Zheng, Yu and Hu, Jie and Xiong, Jiaqi and Liu, Ruiping and Zheng, Junwei and Yang, Kailun and Zhang, Jiaming},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026},
  url = {https://arxiv.org/abs/2512.15621}
}

Please also cite the source datasets, corruption operators and upstream models used in your evaluation. Source references are listed in FORMAT.md.

For dataset issues, open a discussion with the dataset, protocol path, revision and error message. Do not post tokens or private source-data links.