Unidata opens 45-station network to capture humanoid robot training data

Unidata has launched a 45-station data-capture network across nine active sites to collect synchronized depth, motion and tactile data for humanoid robots. The system is designed to fill a major gap in public robotics datasets as robots move from labs into real-world environments.

Why it matters: - Humanoid robots need training data that includes depth, body motion and physical contact, not just flat video. - Public datasets often miss the signals that matter most when robots interact with real objects in ordinary rooms. - Unidata’s setup is aimed at producing that harder-to-collect data at scale for robotics developers.

What happened: - Unidata built a network of capture stations for egocentric robotics data. - The company now operates 45 stations across nine active sites. - The stations record synchronized multi-stream egocentric data for humanoid-robot developers. - The network runs across three daily shifts. - Unidata says the operation records about 400 hours of data per day.

The details: - Each station is a self-contained rig built to capture depth, full-body motion, hand pose and tactile signal in real interiors. - The system centers on the Pico 4 Ultra headset and its stereo camera. - The stereo camera provides a depth map, metric distance to objects, per-frame camera position and lens parameters for scene reconstruction. - Motion trackers on the hands, feet and waist build a real-time full-body skeleton. - A monocular camera on each wrist helps recover hand pose when an object blocks built-in tracking. - Unidata is also implementing tactile gloves that record pressure at contact points. - The tactile layer is designed to capture grip and load that video cannot show. - The company says building the tactile layer from the start is unusual because many teams add it only after models fail on real objects. - The network has logged around 200 manipulation scenarios. - Those scenarios include grasping, transfer and handling objects with different mass and texture. - Typical scenarios use 10 to 15 props chosen to change how the hand makes contact. - Episodes are capped at 20 minutes to reduce tracking drift. - Automated scripts check every take for frame-rate stability and stream synchronization before delivery.

Between the lines: - The launch reflects a broader shift in humanoid robotics from controlled lab demos toward everyday use. - The biggest bottleneck is no longer only model design. The bottleneck is high-quality embodied data with contact and motion signals aligned in one recording. - Unidata is positioning its robotics practice around multimodal dataset production, not just standard video capture.

What's next: - Unidata is continuing to expand its tactile data layer as part of the capture system. - The company is expected to keep generating scenario-based datasets for humanoid-robot training as the network scales. - More manipulation cases and more recorded hours could follow as developers demand higher-fidelity training data.

The bottom line: - Unidata is betting that humanoid robotics will be limited less by algorithms than by the quality of real-world training data.