nuScenes Dataset: Contents, Schema, Extensions and Access

nuScenes is a public autonomous-driving dataset released by nuTonomy (now Motional): 1,000 driving scenes of 20 seconds each, recorded with a full vehicle sensor suite of 6 cameras, 5 radars and 1 lidar with 360-degree coverage, and annotated with 3D bounding boxes for 23 object classes and 8 attributes. It is free for academic use and available under a commercial licence, and it is downloaded after creating an account and agreeing to its Terms of Use.

Definition and scope

The dataset is described in Caesar et al., “nuScenes: A multimodal dataset for autonomous driving” (arXiv 1903.11027, presented at CVPR 2020). The authors position it as the first dataset to carry the full autonomous-vehicle sensor suite, and as a response to a gap they identify: image-only benchmarks drove progress in detection, tracking and segmentation, but most autonomous vehicles also carry range sensors such as lidar and radar, so methods need to be trained and evaluated on data that contains both. The paper also defines new 3D detection and tracking metrics and reports lidar-based and image-based baselines.

“nuScenes” now names a family, not one archive. The core release is the 3D multimodal dataset; nuImages is a separate image dataset with the same sensor setup; and several expansions add semantic labels, maps, vehicle bus data and prediction tooling. A December 2025 survey by the nuScenes team (“nuScenes Revisited”, arXiv 2512.02448) describes the dataset as the first to include radar data, to feature urban driving scenes from two continents and to be collected with a fully autonomous vehicle operating on public roads. This page treats those descriptions as the authors’ own account.

Sensors, scenes and scale

The figures below are the ones the paper’s abstract states. They are the numbers to quote when a methods section needs to describe the data.

• Sensors: 6 cameras, 5 radars and 1 lidar, all with full 360-degree field of view.
• Scenes: 1,000 scenes, each 20 seconds long.
• Annotation: 3D bounding boxes for 23 classes and 8 attributes.
• Relative scale: the authors state 7 times as many annotations and 100 times as many images as KITTI.

In the database, a scene is a 20-second sequence of consecutive frames extracted from a log, and a sample is an annotated keyframe at 2 Hz, collected at approximately the same timestamp as a single lidar sweep. The ego pose comes from a lidar map-based localisation algorithm; it is two-dimensional in the x-y plane, so its z value is always 0. Scenes come from several logged locations, for example the log table’s location field holds values such as singapore-onenorth, and the devkit’s known-issues list also names singapore-hollandvillage, singapore-queenstown and boston-seaport.

What is in the download

The devkit README is explicit about layout. To download nuScenes you create an account on the download page and agree to the Terms of Use; after logging in you see several archives, and the devkit needs all of them. They are unpacked into /data/sets/nuscenes (or any folder, set through the dataroot parameter of the NuScenes class) without overwriting folders that occur in more than one archive. The result has four parts:

• samples: sensor data for keyframes.
• sweeps: sensor data for intermediate frames.
• maps: rasterised .png images and vectorised .json files.
• v1.0-*: JSON tables with all metadata and annotations; each split (trainval, test, mini) has its own folder.

The practical consequence is that a keyframe-only workflow (annotated samples) and a dense-temporal workflow (sweeps included) read from the same tree. nuImages is different: for the nuImages devkit you need at least the metadata and the samples, and the sweeps are optional.

Installing the devkit

The devkit is installed with pip install nuscenes-devkit. The README’s changelog lists v1.2.0 of 28 August 2025 as moving support to Python 3.9 and Python 3.12. The devkit code is released under the Apache 2.0 licence with a 2021 Motional copyright; that licence covers the code, not the dataset, which has its own Terms of Use.

The relational schema

All annotations and metadata, including calibration, maps and vehicle coordinates, are stored as a relational database. Every row has a unique primary key called token, and foreign keys such as sample_token point to the token of another table. For a researcher writing data-loading code or a data-management plan, these tables carry most of the semantics:

• scene, sample and sample_data. A scene is a 20-second sequence; a sample is an annotated keyframe at 2 Hz; sample_data holds an individual sensor record (image, point cloud or radar return) with an is_key_frame flag. Non-keyframe records point to the sample that follows closest in time.
• sample_annotation. A 3D box with location (centre x, y, z), size (width, length, height), a quaternion rotation, and counts of lidar and radar points inside the box. Annotations attach to the sample, not to a single sample_data row, because annotators take all relevant sensor data into account. Attributes live here because they can change over time.
• instance and category. An instance is a particular observed object; categories form a taxonomy with dotted names such as human.pedestrian.adult. Instances are not tracked across scenes.
• calibrated_sensor and sensor. Extrinsics are given relative to the ego vehicle body frame, and camera images come undistorted and rectified. Sensor modality is one of camera, lidar or radar.
• visibility. The fraction of an annotation visible across all six images, binned into 0-40%, 40-60%, 60-80% and 80-100%.
• log and map. The log records vehicle name, capture date and location; the map table holds binary top-down semantic masks (the original release’s semantic_prior for drivable surface and sidewalk).

Extensions: nuImages, lidarseg, Panoptic nuScenes and more

The extensions add labels or context to the core data. The data table below lists each one with its release month, what it adds and where the README says to place it. Two practical points follow from it. First, lidarseg and Panoptic nuScenes both cover the approximately 40,000 keyframes in nuScenes, and both extract a folder together with the v1.0-* tables, so they must be installed on top of the base tree. Second, the map expansion has its own version history: v1.0 (July 2019, 11 semantic layers), v1.1 (resolved ego poses off the drivable surface), v1.2 (arcline paths and lane connectivity for the prediction challenge) and v1.3 (a BitMap class supporting the lidar basemap and the legacy semantic prior map, and one broken lane removed). Cite the map version you used.

Tasks nuScenes supports

The README and the paper together name these tasks and benchmarks: 3D object detection and tracking (the original tasks, with metrics defined in the paper); lidar semantic segmentation (nuScenes-lidarseg); panoptic segmentation and panoptic tracking (Panoptic nuScenes, with metrics such as PAT added to the devkit in September 2021); and trajectory prediction (the prediction challenge, which supplies trajectory sets for the CoverNet approach). The tracking evaluation code was released in November 2019 and the detection evaluation code was reorganised at the same time. For reproducibility, the practical rule is to report the devkit version, the dataset split (mini, trainval or test), the map version and any extension used.

Terms of use and licensing for research offices

Access is gated: the README states that you must create an account on the download page and agree to the nuScenes Terms of Use before downloading. Motional, the company that publishes the dataset, has stated in its announcements that nuScenes is free of charge for academic use and that licensing is available for commercial purposes, with commercial enquiries directed to [email protected]. The December 2025 survey by the nuScenes team describes the terms as free for academics and startups with commercial licences for larger industry partners. Because those two descriptions differ in who is covered, a project that is not plainly academic should read the current Terms of Use on nuscenes.org and record the date it did so.

For a research office, the useful routine is the one used for any restricted-access dataset: keep a copy of the terms accepted, note who accepted them, and record the dataset version in the data-management plan. The site’s guides on data use agreements, versioning research datasets and how to cite a dataset cover those steps in general terms; this page does not substitute for the terms themselves.

nuScenes compared with KITTI

The nuScenes paper uses KITTI as its reference point. The one quantitative comparison the authors give is scale: 7 times the annotations and 100 times the images. The more structural difference, as the paper frames it, is sensor coverage: nuScenes carries 6 cameras, 5 radars and 1 lidar with 360-degree coverage, which makes radar part of the benchmark. For KITTI’s own contents, sensors and licence, see the CASRAI guide to the KITTI dataset. For another widely used benchmark in a different modality, see the guides to ImageNet and MS COCO.

Common misunderstandings

“nuScenes is only the lidar data”

No. The core release is multimodal: cameras, radars and lidar, plus maps and a relational metadata database. Lidar segmentation labels are a separate extension (nuScenes-lidarseg).

“The sweeps folder is optional”

For nuScenes the README says all archives are needed for the devkit to work. Only for nuImages are the sweeps optional.

“Annotations are tracked across scenes”

The schema says instances are not tracked across scenes; tracking is within a scene.

“The devkit licence covers the data”

The Apache 2.0 licence in the repository covers the code. Using the dataset requires accepting the Terms of Use.

“All ego poses are perfect”

The README lists known issues: for boston-seaport the ego poses of scenes 499, 515 and 517 are slightly incorrect and scenes 501 and 502 are outside the annotated area; for the two Singapore locations named above the traffic-light 3D poses are all 0 apart from tz; some lanes are disconnected from the rest; and a small number of 3D boxes are annotated even though the object was temporarily occluded, which is why the nuScenes benchmarks filter objects without lidar or radar points.

Frequently asked questions

What is nuScenes?

A multimodal autonomous-driving dataset of 1,000 twenty-second scenes with 6 cameras, 5 radars and 1 lidar, annotated with 3D boxes for 23 classes and 8 attributes (Caesar et al., CVPR 2020).

Who publishes nuScenes?

It was created by nuTonomy; the devkit repository and licence carry the Motional name, and Motional’s announcements describe the dataset’s terms.

Is nuScenes free to use?

Motional states it is free for academic use with commercial licensing available. Download still requires an account and agreement to the Terms of Use.

How do I install the devkit?

Run pip install nuscenes-devkit. Version 1.2.0 (28 August 2025) supports Python 3.9 and 3.12.

What is the difference between nuScenes and nuImages?

nuImages is a stand-alone image dataset using the same sensor setup and the same devkit. nuScenes is the 3D multimodal dataset.

How should I cite it?

Cite the nuScenes paper (Caesar et al., 2020, arXiv 1903.11027), as the devkit README requests, plus the extension and map version used.

Sources

• Caesar et al., nuScenes: A multimodal dataset for autonomous driving, arxiv.org/abs/1903.11027, read 2026-11-12.
• nuScenes devkit README, github.com/nutonomy/nuscenes-devkit, read 2026-11-12.
• nuScenes database schema, docs/schema_nuscenes.md, read 2026-11-12.
• nuScenes Revisited: Progress and Challenges in Autonomous Driving, arxiv.org/abs/2512.02448, read 2026-11-12.
• Motional, Motional expands groundbreaking nuScenes datasets, read 2026-11-12 (search-result summary).