From:-- Date: 11/19/2026
Nexdata has published five new datasets for Physical AI research. The three egocentric datasets comprise six-camera recordings, monocular video, and 4K stereo video combined with full-body motion capture. There is also a UMI dataset featuring manipulation demonstrations in real homes and a 3D dataset for simulation. Together, the datasets will help researchers study vision-language-action (VLA) models, world models, and robotics.
Latest-generation six-camera setups capture high-resolution multimodal data with wide fields of view and high frame rates.
Highly skilled operators with a lot of hands-on experience performed tasks in 24 real-world scenarios.
The distribution of skills and scenarios was carefully designed and validated based on the specifics of each environment.
20,000 hours of first-person video taken in homes, retail stores, and offices. The tasks range from meal preparation and cleaning to garment care, arranging merchandise, and picking up items.
A PICO 4 Ultra headset records 4K stereo video at 30 fps, while IMUs worn on the wrists and ankles record motion.
The data package includes camera calibration parameters and 76-point full-body pose data (24 body joints and 52 hand joints) time-aligned with the video.
Detailed step-by-step action annotations will help with robot learning and the development of manipulation policies.
135,000 hours of first-person video taken in four different environments – homes, offices, retail stores, and industrial workstations. The tasks involve such activities as organizing household items, binding documents, restocking shelves, and assembling components.
GoPro HERO10 through HERO13 cameras shot 10- to 30-minute videos at 1080p and 60 fps, with hands visible in more than 90% of frames.
Each clip comes with nine-axis IMU data, per-frame camera poses relative to the first frame, relative depth maps, and 3D keypoints for both hands. All data streams are time-aligned.
Subtask annotations indicate the action performed, the skill used, and the start and end times.
50,000 hours of manipulation demonstrations performed in real homes, with about 934,000 episodes and 19.2 million subtasks.
8 scene types, 13 object categories, and 30 action types for household environments and manipulation tasks.
Six synchronized views provided by head- and wrist-mounted cameras for view ablation studies and stereo depth estimation.
Video frames and proprioceptive data aligned at 30 Hz for imitation learning.
Task and subtask annotations with an average of 20 segments per task.
Labels for scenes, objects, and actions for evaluating generalization across environments, objects, and actions.
60 million real-world samples and 40 million synthetic samples.
Some objects have physical properties and motion constraints for interactive components, allowing simulation in NVIDIA Isaac Sim.
3D scenes include residential, retail, office, industrial, hospitality, and tourism environments. Objects include furniture, home appliances, and kitchenware.
Supported formats include OBJ, FBX, glTF, USD, and MAX. Assets come with materials and textures, multi-view images, and structured semantic descriptions.
Please contact us for dataset samples, technical specifications, and licensing information. For custom projects, you can also request bespoke data collection and annotation, including task design and equipment installation.