PhysicalAI Autonomous Vehicles: 1,100 NCore Clips with Cuboid Labels

NVIDIA's PhysicalAI Autonomous Vehicles NCore dataset consists of approximately 1,100 multi-sensor driving clips extracted from a larger 1,700-hour collection. This subset, updated in March 2026, provides high-fidelity data specifically formatted for the NCore standard with verified calibration and labeling.

Use Cases

• Training 3D object detection models using cuboid labels
• Developing vehicle localization algorithms using egomotion data
• Validating sensor fusion pipelines using offline calibration parameters

Strengths

• 1,100 high-quality clips
• Verified offline calibration data
• Includes 3D cuboid labels
• Standardized NCore format

Limitations

• Small sample size relative to the source dataset
• Restricted to clips with specific label types
• License is listed as 'other' which may restrict commercial use

Provenance

Source

NVIDIA

Collection Method

Sensor data collected from autonomous vehicle fleets

Freshness

Last updated March 2026.

Geography

25 countries