0 to $8M+
Egocentric depth video from real plumbers, electricians, and tradespeople - captured in uncontrolled environments, at union scale - is the highest-value training signal for humanoid dexterity. No simulation can replace it. No lab can source it. We build the supply chain.
Real hands, real work, real data
The race to ship general-purpose humanoids is a data problem. Every lab - xAI, Physical Intelligence, Tesla, Boston Dynamics - can build the hardware. None of them can generate enough real, varied, labeled dexterity data from the physical world.
Simulation transfers poorly to the mess of real manual work: tight crawlspaces, wet pipe joints, unlabeled breaker panels, tools that shift in the hand. The only way to train hands that work like hands is to watch real hands working - thousands of them, across every trade, every city, every season. Depth data from an egocentric camera captures what no lab or synth dataset can: the true distribution of human manipulation in the wild.
We are building the only pipeline that can deliver this data at scale. Union partnerships guarantee a rotating workforce of experienced tradespeople. Egocentric dual-camera rigs capture depth and RGB from the worker's perspective. Automated pipelines clean, calibrate, and label the output. Robotics companies buy it by the terabyte.
The first dollar goes to hardware and trust. We source 200 dual-camera egocentric rigs - modified depth sensors paired with compact RGB modules - and recruit the first cohort of 50 union plumbers and electricians across three metro markets. Liability and data-rights agreements are negotiated directly with local union halls, establishing the template that scales to every new market. A manual QA pipeline handles timestamp alignment and basic calibration; automation comes in Phase 2.
With the pilot template proven, we automate everything that was manual. Sensor fusion, occlusion handling, and photometric calibration move from per-clip human review to a pipeline that processes raw rig footage overnight. A semi-automated tagging platform starts with a pool of manual taggers and iterates toward ML-assisted labeling, cutting cost-per-hour with every release. The fleet scales to 500 rigs across 10 metro markets, supported by a dedicated operations lead and three regional coordinators.
By month 11 we have the only signed, repeatable pipeline for egocentric manual labor data at scale. Union partnerships create a structural barrier - labor relationships take years to build, and no competitor can replicate them on a fast timeline. Depth data from real manual work is the highest-value training signal for dexterity, and every major robotics lab - xAI, Physical Intelligence, Tesla - needs it. None of them can source it themselves; they buy it from us.