Teleoperation as Human Labeling

This makes teleoperation a training input, not the end product. The core bet is that every time a remote expert steps in, 1X gets another labeled example of what the robot saw, what action solved the problem, and what happened next. That data then feeds Redwood and 1XWM, so the robot can handle the same kind of situation by itself later, which is how human oversight cost can fall as deployments rise.1 2 3 4 5

• 1X describes its data recipe as a mix of web video, egocentric human video, simulation, remote operator robot data, and on policy robot data. That framing puts teleoperation in the same role as annotation in AI training, useful because it produces high quality examples in the exact situations the robot fails today.3 4
• The product already shows this bridge model in practice. NEO is sold with basic autonomy, plus Scheduled Expert Mode for tasks it does not yet know, and a separate remote control mode. That means teleoperation helps the robot complete work now, while also generating the examples needed to automate those tasks later.4 6 7
• This is different from robots built around permanent remote operation. In 1X and Figure, human intervention is part of the learning loop. In other settings, like hazardous work, teleoperation can itself be the product. 1X is aiming for the first model, where autonomy expands and the operator recedes into edge cases.5 8 9

If 1X keeps increasing robot deployments, the constraint shifts from finding teleoperators to turning intervention logs into reusable autonomy faster than rivals. The winning home robot company is likely to be the one that converts each supervised chore into a capability that thousands of later robots can perform alone.2 3 5