Awesome-WAMs
GitHub
RESEARCH NOTEref-236eff7a0d0d26ae758etechnical resource

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving

Catalog authors & publication details

OpenScene Contributors

View the full catalog record
Resource reviewed5 min readPrepared 9 Sept 2026

OpenScene is a compact redistribution of nuPlan with additional occupancy labels and occupancy-grid motion annotations. The supplied README describes the data resource and its challenge uses; it does not specify or evaluate a learned world-action model. This report covers that overview only. The reported scale and storage reduction describe dataset packaging, not measured prediction or driving performance.

Open primary source Inspect reading coverage
Resource overview

This note reviews a research resource. Its scope is recorded explicitly and is separate from a full-paper review.

01

The idea

The problem

Source description

The resource addresses the burden of using a large autonomous-driving dataset for occupancy research. It retains relevant nuPlan annotations and sensor data at 2 Hz and adds occupancy labels across Boston, Pittsburgh, Las Vegas and Singapore. The README motivates compact access to broad driving coverage, without defining a particular prediction horizon or evaluation protocol. e-data

What this work contributes

Author claim

The maintainers report a dataset-size reduction by a factor greater than 10 while retaining over 120 hours of coverage. No byte counts or measurement procedure accompany the reduction claim. e-data

Source description

The statistics table lists OpenScene as nuPlan-based, with 120 sensor-data hours and flow annotations. Its semantic-category entry remains TODO. The prose says over 120 hours, so the tabular 120 should not be presented as an exact reconciled total. e-datae-statistics

Source description

The README identifies OpenScene as data used for the End-to-End Driving and Predictive World Model tracks at the CVPR 2024 Autonomous Grand Challenge and the NAVSIM-v2 End-to-End Driving track at the 2025 challenge. These are documented use cases, without accompanying performance results. e-challenges

How to read the labels

Source description summarizes the inspected material. Author claim preserves the authors’ attribution. Reader analysis and Open question are interpretive.

02

Mechanism & design

INPUTS
  • Dataset construction: relevant annotations and sensor data from nuPlan
OUTPUTS
  • A compact redistribution sampled at 2 Hz
  • Additional occupancy labels and per-grid motion direction and velocity annotations
  1. 01

    Retain a compact driving-data subset

    The resource description specifies retaining relevant annotations and sensor data at 2 Hz. It does not identify the complete retained sensor inventory, the selection rule for relevance, or filtering criteria. e-data

  2. 02

    Construct temporally accumulated occupancy annotations

    The stated span of LiDAR frames accumulated for each occupancy annotation is 20 seconds. This is an annotation-construction span; the README does not identify it as a model input duration or future prediction horizon. Registration, motion compensation and temporal alignment details are absent from this overview. e-annotations

  3. 03

    Attach motion information to occupancy grids

    The source defines flow as motion direction and velocity for each occupancy grid. It does not specify units, coordinate frames, validity masks or an annotation-generation algorithm. Getting Started points to separate Download Data and Prepare Dataset documentation. e-annotationse-getting-started

03

Taxonomy assessment

Catalog at reading time

Catalog updated

Major category
Datasets
Subcategories
Autonomous driving perception data · 3D occupancy & scene flow data · Benchmarks & evaluation protocols
Architecture
Not applicable
Prediction paradigm
Not applicable
Quadrant
Not applicable
Classification status
Explicit in survey
READING ASSESSMENT

Supports the recorded classification

Reader analysis

The snapshot supports the recorded dataset category and its autonomous-driving occupancy and scene-flow emphasis. Challenge use supports a benchmark role, but the detailed evaluation protocols remain unverified. Architecture, prediction paradigm and quadrant are appropriately Not applicable for this resource overview: it defines neither a unified world/action predictor nor an inverse-dynamics control mechanism. e-datae-statisticse-annotationse-challenges

These labels preserve the catalog snapshot used for this reading. The assessment audits that snapshot without changing the source classification.

04

Limits & reproduction

Limitations and open boundaries

Reader analysis

The comparison table describes dataset hours, flow availability and semantic-category counts. It provides no model accuracy, baseline training protocol, train/test split, uncertainty, ablation or executed-driving success. Its entries cannot establish that OpenScene produces better perception or control. e-statisticse-challenges

Reader analysis

The semantic-category count for OpenScene is explicitly unfinished in the table. The overview also leaves occupancy resolution, spatial extent, class definitions, sampling balance and split construction unspecified. These omissions limit assessment of label compatibility, geographical balance and leakage risk. e-statisticse-datae-annotations

Open question

Twenty-second accumulation makes the temporal alignment of supervision a relevant unresolved issue. The overview does not say whether the window uses frames before, after or around a target timestamp, or how moving objects are handled. It therefore neither establishes leakage nor demonstrates its absence. e-annotations

Source description

The README distinguishes data and code licensing: the dataset is distributed under Creative Commons Attribution-NonCommercial-ShareAlike and the nuPlan Dataset License Agreement for Non-Commercial Use; repository code is described as Apache License 2.0. The README does not specify a Creative Commons version. e-license

What a reproduction would require

Reader analysis

Reusing this resource requires the separately linked download and preparation instructions plus the applicable nuPlan and OpenScene data terms. This overview alone cannot define a reproducible benchmark: an immutable data release, split lists, label specification, evaluator and any baseline configuration still need to be established. e-getting-startede-licensee-statistics

Reader analysis

Proposed check, not performed: once a fixed release and its preparation rules are available, compare retained timestamps and file sizes with the corresponding nuPlan subset. Hold scenes and storage accounting constant; test whether retained samples follow the stated 2 Hz cadence and whether the claimed greater-than-tenfold reduction holds. A mismatch would challenge that packaging claim under the recorded accounting, without implying anything about model quality. e-data

Reader analysis

Proposed check, not performed: for fixed moving-object scenes, reconstruct occupancy using the documented 20-second window and a single-frame control, with identical grids and timestamps. Inspect whether accumulated labels smear moving objects and compare flow direction with independently tracked displacement in the documented frame. Report any use of future observations separately. Persistent disagreement would question temporal label consistency; it would not by itself measure driving performance. e-annotations

05

Questions to take further

  1. What immutable release and split definition correspond to the README statistics and each challenge use?
  2. How are the 20-second accumulation window, moving-object compensation and flow reference frame defined?
  3. What semantic labels replace the table’s TODO, and how are label coverage and geographical balance evaluated?
06

What was read

Resource overviewSource identity verified
Sections inspected
  • Complete supplied repository landing-page text, including navigation and footer
  • OpenScene: Autonomous Grand Challenge Toolkits
  • Description, including dataset statistics and annotation notes
  • Getting Started
  • License and Citation, including all three BibTeX entries
  • Related Resources
  • About
Appendix
Not present
Figures inspected
No figures recorded as inspected
Tables inspected
Description: unnumbered dataset statistics table, extracted text only

Outside this reading

  • External figure assets are not downloaded; HTML alt text and mathematical text are retained where available.
  • Only the repository landing README is retained; linked dataset preparation/download documentation, code and datasets were not read.
  • The complete supplied HTML text was read, but the acquisition label full-text refers to a repository overview, not a complete research paper or complete resource documentation.
  • No usable primary PDF or original visual assets were supplied. The dataset table was read as extracted text only; no figures, table images or title-page image were visually inspected. No illustrated edition is created.
  • Identity/version note: the README heading is OpenScene: Autonomous Grand Challenge Toolkits. Its openscene2023 citation gives the exact catalog title, OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving, author OpenScene Contributors, and year 2023. The catalog uses citation key openscene. The citation verifies the resource identity despite the different current heading.
  • The reviewed artifact is the repository documentation snapshot accessed 2026-09-07T14:30:41.048810+00:00. No immutable commit or numbered release is established, and the original 2023 documentation was not recovered or compared. The current snapshot also mentions 2024 and 2025 challenge uses.
  • The separately cited Visual Point Cloud Forecasting enables Scalable Autonomous Driving and Scene as Occupancy papers were not supplied or reviewed. Their authors and methods are not attributed to this resource.
  • Code, datasets, linked challenge documentation and full license texts were not inspected; no experiments or dataset checks were run. Affiliations are not stated in the retained resource citation. HTML heading and citation locations replace a PDF page in the title metadata.
07

Evidence & sources

Evidence links resolve to these source locations. Expand an entry to inspect its supporting detail.

e-identityREADME heading OpenScene: Autonomous Grand Challenge Toolkits; License and Citation, @misc{openscene2023}Inspect

The resource heading differs from the citation title. The OpenScene BibTeX entry gives the exact catalog title, OpenScene Contributors, year 2023 and the OpenDriveLab/OpenScene repository URL. Separate BibTeX entries credit other works.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving
e-dataDescription, opening paragraphInspect

The README describes a compact nuPlan redistribution retaining relevant annotations and sensor data at 2 Hz, claims a size reduction by a factor greater than 10 and over 120 hours of coverage, and names Boston, Pittsburgh, Las Vegas and Singapore for the additional occupancy labels.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving
e-challengesDescription, paragraph beginning OpenScene is also the large-scale datasetInspect

The README lists the CVPR 2024 challenge End-to-End Driving and Predictive World Model tracks and the CVPR 2025 NAVSIM-v2 End-to-End Driving track, referring readers to challenge documentation for details.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving
e-statisticsDescription, unnumbered dataset statistics table; OpenScene row and column headersInspect

The columns are Dataset, Original Database, Sensor Data (hr), Flow and Semantic Categories. OpenScene is listed with nuPlan, 120 hours, a flow check mark and TODO for semantic categories. The table reports resource statistics, not predictor scores.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving
e-annotationsDescription, two notes immediately following the dataset statistics tableInspect

Each occupancy annotation accumulates LiDAR frames spanning 20 seconds. Flow is defined as motion direction and velocity for each occupancy grid. These notes provide no further temporal alignment or flow-generation specification.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving
e-getting-startedGetting Started, complete sectionInspect

The retained section contains Download Data and Prepare Dataset references; their linked instructions are not present in the supplied overview.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving
e-licenseLicense and Citation, first two paragraphsInspect

The README states that the dataset follows Creative Commons Attribution-NonCommercial-ShareAlike and the nuPlan noncommercial dataset agreement, requires attribution and excludes commercial use. It describes repository code as Apache License 2.0.

OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving

Source record