Teaching robots how to build.

Construction is one of the hardest environments for physical AI. Every project is different, the workspace changes continuously, and the thing a worker is trying to create often does not exist yet. Much of robotics today is trained from egocentric video, teleoperation and human demonstrations that capture what people see and what they do. But in construction, that is only part of the story. Every action is shaped by something outside the scene: the design, the target state, the sequence, the thing that should exist next.

OJOx started from a simple question: what if we captured that too? We bring design intent directly into the worker’s field of view, registered to the physical workspace, so they can see what needs to happen in place and at scale. As they work, we capture the experience around that intent: what they see, how they move, how they use their hands, the surrounding space, and how the task changes over time. The goal is not simply to collect more demonstrations. It is to collect the right demonstrations, where what should be built and how a skilled person builds it are recorded together.

I think this creates a very different path for robotics in construction. The same interface can help guide people through complex physical work today, while capturing the knowledge needed to teach machines tomorrow. Rather than separating human assistance from robot learning, OJOx treats them as part of the same loop: guide the human, capture the skill, teach the machine. Each task completed by a person can become a better demonstration of how intent becomes physical reality.

That is the direction we are pursuing with OJOx. We want to build the data layer for construction robotics, starting with the people who already know how to build. And over time, we want to answer a much bigger question: can a machine be shown what should exist, even when it has never built it before, and use learned physical skill to make it real?

Guide the human. Capture the skill. Teach the machine.