nuScenes

nuScenes is a multimodal autonomous-driving dataset released by Motional (formerly nuTonomy). It comprises 1,000 driving scenes of about 20 seconds each, recorded in Boston and Singapore and hand-selected for dense traffic, rain, night driving, and challenging urban maneuvers. Each vehicle carried a full sensor suite, giving researchers time-synchronized, 360-degree coverage of complex real-world traffic, along with detailed map layers for context.

What’s in it

The sensor rig includes 6 cameras, 1 spinning 32-beam LiDAR, and 5 radars, complemented by GPS and IMU ego-pose. Annotations cover roughly 1.4 million 3D bounding boxes across 23 object classes with attributes and tracking IDs, sampled at 2 Hz keyframes, totaling about 15 hours of driving. Related releases such as nuScenes-lidarseg and nuImages add point-level and image segmentation labels, and a prediction split targets trajectory forecasting. The release also bundles vehicle CAN-bus signals and a rasterized semantic map with drivable areas, lanes, and crosswalks, supporting map-aware perception and planning.

Tasks, license & access

nuScenes underpins benchmarks for 3D object detection, multi-object tracking, LiDAR segmentation, and motion prediction, and it popularized the nuScenes Detection Score (NDS) evaluation metric. The data is free to download after registration but is licensed strictly for non-commercial use (CC BY-NC-SA 4.0), with a separate commercial license available through Motional.

• Modality: camera, LiDAR, radar, ego-pose (sensor fusion)
• Scale: 1,000 scenes, ~1.4M 3D boxes, 23 classes, ~15 hours
• Tasks: 3D detection, tracking, LiDAR segmentation, prediction
• License: CC BY-NC-SA 4.0 (non-commercial)
• Creator: Motional (nuTonomy/Aptiv)

As one of the first full-sensor-suite AV datasets, nuScenes remains a standard benchmark for multi-sensor perception in physical-AI systems.