OccStress

Stress-Testing the 4D Occupancy Forecasting Chain

Paper | Dataset | External Data Setup | Licenses and Use

OccStress evaluates how errors in observed 3D occupancy states affect future occupancy forecasts. It pairs sensor-corrupted upstream predictions with controlled occupancy-state interventions and temporal diagnostics, supporting source-specific and corruption-specific robustness analysis.

This repository contains protocol definitions, derived occupancy assets and supporting metadata for OccStress-nuScenes, OccStress-Waymo and OccStress-CARLA. It does not redistribute original camera images, LiDAR point clouds, official clean GT or model checkpoints.

Availability, 2026-10-06: Data preparation is complete. Package upload is in progress. Consult packages.json for the current status: only packages marked available have passed remote size/content checks. The overall status becomes complete after every planned package is verified.

Benchmark Tracks

Track	What is provided	Evaluation purpose
Manual	Semantic replacement, dropout, holes, misalignment and traffic mirroring	Measure sensitivity to controlled occupancy-state interventions
Upstream	Occupancy predictions from 3D models under camera or LiDAR stressors, with source-specific clean references	Measure errors propagated through perception into forecasting
Position sweep	A clean reference and five single-state injection positions per setting	Isolate input-position effects at fixed corruption severity and budget

The main Current-only, Recent-burst and History-only regimes combine temporal position and duration. They measure scenario-level robustness, not isolated position effects. Traffic mirroring is a coordinate-consistent layout transfer and is reported separately from non-traffic manual robustness.

Dataset Coverage

Dataset	Domain	Scenes	Frames	Anchors per protocol	Manual	Upstream	Position sweep
OccStress-nuScenes	Real-world, Occ3D-nuScenes validation	150	6,019	4,519	38	364	60
OccStress-Waymo	Real-world, Occ3D-Waymo validation at 2 Hz	202	7,998	5,978	38	219	0
OccStress-CARLA	Simulated, UniOcc CARLA validation at 2 Hz	3	360	330	38	218	0

The prepared release contains 975 protocols, 326 asset conditions and 1,654,143 occupancy/event files, totaling approximately 136.9 GB (127.5 GiB) before archive packaging. Metadata, external datasets, checkpoints and temporary extraction space are additional.

Protocol counts include their clean references. Protocol records reuse anchors and assets; they are not independent scenes or unique samples. Waymo's two EFFOcc sensor-stress branches retain separate clean protocol files.

The nuScenes position sweep has ten hard settings, each with one clean reference and five corrupted positions. Cross-model comparisons should use the four jointly observed positions when a model does not observe the earliest state. An unchanged result for an unobserved state is not evidence of robustness.

Download and Setup

The download entry point is insailab/OccStress. The package index lists archive paths, availability, file counts, unpacked sizes and archive SHA256 values. Each archive is independent, not a split volume; extract selected archives into the same root. Asset shards target 2 GiB before compression and are grouped by dataset, track and upstream source.

To download all currently published files:

python -m pip install huggingface_hub
hf download insailab/OccStress --repo-type dataset \
  --local-dir /datasets/OccStress-download

For a smaller download, this example selects Waymo manual assets and the corresponding protocols/metadata. Check that all required shards are available before evaluation; an in-progress upload is not a complete benchmark download.

hf download insailab/OccStress --repo-type dataset \
  --local-dir /datasets/OccStress-download \
  --include README.md EXTERNAL_DATA.md DISTRIBUTION.md packages.json \
  'manifests/*' 'metadata/common.tar.zst' 'metadata/OccStress-Waymo.tar.zst' \
  'archives/manual/OccStress-Waymo/*.tar.zst'

Metadata packages contain all protocols for their dataset. Running an upstream or position-sweep protocol additionally requires the corresponding asset shards. Position sweeps reuse manual/upstream assets; they have no duplicate voxel packs.

Downloading files does not extract archives or obtain external GT. Use a separate directory for original dependencies:

export OCCSTRESS_DATA_ROOT=/datasets/OccStress
export OCCSTRESS_EXTERNAL_ROOT=/datasets/OccStress-external
mkdir -p "$OCCSTRESS_DATA_ROOT"
find /datasets/OccStress-download/metadata /datasets/OccStress-download/archives \
  -type f -name '*.tar.zst' -print0 | \
  while IFS= read -r -d '' archive; do
    tar --zstd -xf "$archive" --no-same-owner -C "$OCCSTRESS_DATA_ROOT"
  done

The extraction example uses Bash, GNU tar and zstd. The archives contain relative paths directly under meta/, protocols/, occ/ and events/; do not strip a path component. Keep manifests/ and the documentation beside those directories if you also want the release records in the extracted tree. Archive checksums refer to the compressed files, not individual NPZ assets.

In the OccStress code checkout, follow EXTERNAL_DATA.md to mount authorized Occ3D GT and official info files, or prepare CARLA's canonical GT. Then check the dataset you intend to use:

python tools/prepare_external_data.py check --dataset nuscenes --trust-pickle
python tools/prepare_external_data.py check --dataset waymo --trust-pickle
python tools/prepare_external_data.py check --dataset carla --trust-pickle

These are independent commands; only run the checks for datasets you prepared. Model evaluation also requires the selected forecaster's environment and checkpoint. Upstream 3D weights are needed only to regenerate exports, not to consume the provided predictions. Use the model-specific instructions in the code release.

This is a file-based benchmark consumed by OccStress adapters, not an Arrow table to open directly with datasets.load_dataset. Protocol PKLs must come from a trusted source; do not unpickle arbitrary downloaded files.

File Layout

OccStress/
  README.md
  EXTERNAL_DATA.md
  DISTRIBUTION.md
  manifests/
  meta/
    OccStress-nuScenes/
    OccStress-Waymo/
    OccStress-CARLA/
  protocols/
    manual/<dataset>/
    upstream/<dataset>/<subtrack>/<source>/
    position_sweep/OccStress-nuScenes/<setting>/
  occ/
    manual/<dataset>/<family>/<severity-if-applicable>/
    upstream/<dataset>/<subtrack>/<source>/<condition>/
  events/
    manual/<dataset>/

<dataset> is one of OccStress-nuScenes, OccStress-Waymo or OccStress-CARLA. Paths in protocols are relative to this shared root. References beginning with external/<dataset>/ resolve through OCCSTRESS_EXTERNAL_ROOT; external assets are not bundled in this tree.

Location	Contents
meta/<dataset>/dataset.json	Scene/frame/anchor counts, temporal window, grid and source metadata
meta/<dataset>/	Class mapping and dataset-specific pose/control metadata or dependency descriptions
protocols/	Anchor-aligned evaluation records and corruption/temporal settings
occ/	Canonical derived occupancy NPZ assets shared across temporal protocols
events/	Corruption event metadata where provided
manifests/	Inventories and validation coverage

Upstream clean predictions are model outputs, not official clean GT. Misalignment uses transformation metadata rather than a second full occupancy copy. Do not duplicate assets for every temporal protocol.

Temporal and Label Conventions

Field	Convention
Canonical input window	Four historical states plus current: -2.0, -1.5, -1.0, -0.5, 0.0 seconds
Future targets	Six actual future states: +0.5, +1.0, +1.5, +2.0, +2.5, +3.0 seconds
Derived grid	200 x 200 x 16
Canonical labels	Occ3D 18-class space, including free space; actual class support is dataset-dependent
Anchor selection	Complete H4/current/F6 windows within a scene; identical anchors across its protocols

GenieDrive uses the last four observed states and does not observe -2.0 seconds. Record the actual model input window and control strategy in every comparison. Do not replace the first future target with current-frame reconstruction.

Native Waymo voxel_label must be mapped using the supplied class map. Derived semantics arrays are already canonical and must not be mapped again. CARLA's native UniOcc arrays are left-handed; the evaluation view is x-forward/y-left/z-up. The external preparation tool maps labels and flips grid axis 1 exactly once while preserving the published right-handed controls.

For CARLA, command order is [right, left, straight]. At the sixth future state, local y <= -2 m selects right and y >= 2 m selects left; otherwise straight. Non-anchor tail frames retain future-validity masks and a straight fallback. Every evaluated anchor has six valid future targets. Traffic mirrors y consistently in occupancy, poses and trajectories, and swaps the right/left command entries. Preserve these transformations when adapting another model.

Evaluation and Quality Checks

Report semantic mIoU and binary occupancy IoU per future horizon and identify which horizons are averaged. Keep the metric convention, class support and control policy explicit. Historical paper scores and the code release's occstress-present-gt-v1 convention can differ in absent-class handling and must not be mixed without qualification.

Preparation checks cover protocol structure, anchor counts, relative-path resolution, NPZ members, grid shapes, legal labels and finite values. CARLA controls and traffic transforms have separate audits. The complete clean dependency check covers 14,377 GT references for existence and nonzero size. Loader equivalence and GPU compatibility checks are sampled, not exhaustive reruns of every model/protocol combination.

Public coverage is recorded in manifests/protocols.json, manifests/validation_coverage.json and the CARLA control/metadata audit files. Internal job records and condition-level completion markers are not distributed. Event JSON files omit machine-specific source_path and output_path provenance; their numerical corruption metadata and occupancy contents are preserved. Use packages.json, not preparation receipts, to check download availability. Preparation status does not certify distribution permission or paper-score reproduction.

Intended Use and Limitations

• Intended for research on occupancy forecasting robustness, error propagation and controlled temporal sensitivity; not a deployment safety certification.
• Corruptions are reproducible stress conditions, not samples from a calibrated distribution of real-world failures. Scores are not deployment-risk estimates.
• nuScenes and Waymo cover limited real-world driving domains. CARLA has only three simulated validation scenes and cannot substitute for broad real-world validation. Their label support and sensor configurations differ.
• This release supplies evaluation subsets, not new training splits. Report zero-shot evaluation separately from training or tuning on the target dataset.
• Available protocols do not imply that every model supports every track. Camera-direct forecasters and withheld third-party adapters have separate applicability constraints in the code release.

Sources and Terms

The source datasets are Occ3D, nuScenes and Waymo, plus UniOcc CARLA (Carla-2Hz-val, three 120-frame scenes). Cite the corresponding datasets, corruption operators and upstream models in addition to OccStress.

Access to these derived assets does not grant access to the source datasets or override their terms. Obtain the required source registration and permissions before use. This collection has source-dependent terms, not a blanket MIT license; 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}
}

For dataset issues, use the repository discussion page and include the dataset, protocol path, code revision and error message. Do not post access tokens, private source-data links or identifying machine paths.