Physical AI Data

Annotation of first-person footage where the camera moves with the wearer: hand and object tracking, action and sub-action segmentation, hand–object interaction, gaze and attention labels, and temporal event tagging. Identity and action boundaries are held across the full sequence despite constant camera motion, motion blur and frequent occlusion by the wearer’s own hands.

Training data for imitation learning and VLA models: human demonstrations from egocentric and fixed views, teleoperation episode annotations, and key training signals – task boundaries, grasp phases, outcomes, and paired natural-language instructions.

3D cuboids, point cloud segmentation, object tracking across LiDAR sweeps and camera – LiDAR fusion annotation, for autonomous vehicles, mobile robots, drones and mapping. Annotators work in native point cloud formats, hold instance IDs across frames, and reconcile objects across sensors before delivery.

Hand keypoints, 3D hand pose, grasp types and object contact states, tracked frame to frame through the self-occlusion that defines first-person video. Skeletal and landmark annotation also covers full-body pose for gait, ergonomics, sports and assistive-technology models, with joint consistency maintained through partial occlusion.

 Evaluation sets and expert review for real-world systems: edge-case/failure-mode collection, expert-labelled test scenarios, and structured behavioural scoring against ground truth. Test policy generalisation before real-world deployment.

Surgical robotics, endoscopy, ultrasound, and AI data: DICOM annotation, phase/instrument tracking, and clinical video labeling by credentialed specialists with regulatory documentation.