45 Stations, 400 Hours a Day: What a Robot Data Factory Actually Looks Like

Egocentric video data collection is the practice of recording first-person footage — from head-mounted, chest-mounted, or wrist-mounted cameras — used to train robots and embodied AI systems on how humans see, move, and manipulate objects. It's the closest visual proxy available for what a robot's onboard camera will actually see once it's deployed, which is why demand for it has outpaced almost every other category of AI training data this year.

What's new isn't the concept. It's the operational scale. Unidata, a data-infrastructure vendor focused on embodied AI, disclosed that it now operates 45 stations across nine active sites, recording synchronized multi-stream egocentric data for humanoid-robot developers . Running at three daily shifts across the network, the company records on the order of 400 hours of this data per day . That's not a dataset release. That's a facility with a shift schedule.

For anyone running field data operations — ours included — the interesting part isn't the robot angle. It's that egocentric capture has quietly become a logistics problem: staffing, hardware calibration, site rotation, and quality control at volume, not a one-time recording session.

What Changed: From Footage to a Capture Network

Back in July, Unidata announced a capture system built to close a specific gap in robot training data: the absence of spatial and proprioceptive signal in most egocentric footage. Unidata, a data infrastructure company focused on embodied AI, launched a multimodal egocentric capture system for humanoid-robot training datasets . In one session it records synchronized stereo video, full-body skeletal motion, and continuous camera pose — the spatial and proprioceptive signals robots use at inference — plus per-frame depth maps from the stereo footage .

That was the pilot. The more recent disclosure describes what it looks like running continuously. The build centers on the Pico 4 Ultra headset and its stereo camera, which supplies a depth map, metric distance to objects, per-frame camera position, and the lens parameters needed to reconstruct a scene . Motion trackers on the hands, feet, and waist build a real-time full-body skeleton, while a monocular camera on each wrist recovers hand pose when an object blocks the built-in tracking . The Robot Report picked this up too, noting that on the skills-training side, Unidata describes its approach to egocentric data collection as part of a broader look at how mobile manipulators and humanoids are sourcing training data.

The Signals Public Datasets Don't Have

None of this matters much unless you understand what's missing from the datasets everyone already has. The public egocentric corpus is genuinely enormous now. Build AI's Egocentric-1M, released in April 2026, reached one million hours of factory footage, the largest ever assembled . That's not a modest number — it's roughly two orders of magnitude larger than what existed in late 2025.

But volume and usefulness aren't the same axis. Nearly all egocentric video available for training robots is flat 2D: no depth, no camera pose, no record of physical contact. It is fine for pretraining and largely useless for the final stage of policy training, where a robot needs to know exactly how far its hand is from an object, not just that a hand is present in frame. Research on manipulation policies backs this up directly: adding an egocentric wrist-mounted view has been shown to improve task success by more than 50 percent over third-person baselines on tasks requiring fine-grained spatial alignment .

Consider a simple example: a warehouse-picking policy trained only on Ego4D-style footage learns roughly where a hand goes near an object, but not the depth, force, or contact geometry needed to actually close a gripper on it. That's the gap Unidata's stations are built to fill — and it's the same gap our egoaffordance labeling work keeps running into on the annotation side.

Signal	Consumer capture (GoPro/phone)	Purpose-built capture station
RGB video	Yes, 2D only	Yes, stereo
Depth / metric distance	No	Yes, per-frame
Camera pose	No	Yes, continuous 6DoF
Full-body skeleton	No	Yes, via hand/foot/waist trackers
Occluded-hand recovery	No	Yes, wrist-mounted monocular cameras
Tactile / contact force	No	Piloting instrumented gloves

Why This Reads Like a Field Operations Problem, Not a Data Problem

Forty-five stations across nine sites, running three shifts a day, is a staffing and scheduling exercise before it's an AI exercise. Someone has to calibrate 45 rigs, rotate operators, keep the stereo cameras synced, and flag bad takes before they get baked into a dataset. That's field operations vocabulary, not machine learning vocabulary — and it's the same discipline that determines whether a mystery-shopping program or a GIS survey produces usable data or noise.

It also echoes a point we've made about robot data pricing: the cost of a capture hour isn't set by the camera, it's set by the operational overhead around it — site access, scheduling, QA review, re-shoots. Our piece on what a per-hour price actually has to cover goes into this in more depth, but the short version applies here too: a 400-hour daily output number tells you almost nothing about cost or quality without knowing how many of those hours pass QA.

There's also a body-specific dimension worth flagging. Unidata is implementing tactile gloves that record pressure at the contact points, capturing grip and load that video cannot show — because, as the underlying reasoning goes, visual data can't capture force distribution across a gripper, surface texture, or an object's mass asymmetry, since a tool may look uniform yet weigh more on one side . Hands need a different capture rig than legs or torsos, which is the same argument we laid out in our look at why robot data collection isn't one job.

The Equipment Tiers, If You're Scoping This Yourself

If a client asks what egocentric capture requires before committing budget, the honest answer depends on which policy the data will train. Unidata's own technical writeup breaks it into three practical tiers:

• Baseline: a head or chest-mounted camera such as a GoPro, iPhone, or Pico 4 Ultra captures first-person RGB at low hardware complexity — sufficient for coarse action recognition, not for manipulation policies.
• Intermediate: adding a body motion tracker, such as Xsens or a similar system, provides skeleton and wrist orientation data , closing part of the pose gap.
• Advanced: a multi-camera rig with wrist-mounted cameras, a fisheye chest camera, a depth sensor, and a synchronized IMU array matches robot sensor geometry and captures the signals needed for dexterous manipulation learning .

The practical guidance is blunt: match the rig to the policy's actual input requirements, not to whatever hardware looks most impressive in a demo. A locomotion policy doesn't need tactile gloves. A bin-picking policy probably does.

Frequently Asked Questions