1X launches a world model lab for Neo, betting robot data beats fine-tuning

Why it matters

1X is trying to make manufacturing a model advantage: if Neo robots can generate better embodied data, the company may have a moat that software-only AI labs cannot copy quickly.

1X CEO Bernt Bornich is launching the 1X World Model Lab, an internal push to train foundation models for its Neo humanoid robot with robot-native data rather than treating robotics as a late-stage fine-tuning problem, Forbes reported.

Bornich's argument is simple and pointed: "You can't fine-tune your way to AGI," he told Forbes. "You need to actually train the model properly." For 1X, that means making the robot itself, the data it generates and the model training loop part of the same system.

The new lab will be run by Sam Sinha, whom Forbes describes as a founding researcher at video-generation startup Luma AI and 1X's new Head of World Models. Sinha spent the last four years scaling multimodal models at Luma, according to Forbes, a useful credential for a robotics company trying to turn video, physical feedback and robot behavior into a training advantage.

The bet: start with embodied data

The lab builds on the 1X World Model that Forbes says 1X launched in January 2026. That earlier model was trained on video data and let Neo turn prompts into actions, including interactions with objects the robot had not previously seen, according to the report.

The shift now is not just a new research group. It is a claim about where durable advantage in humanoid robotics will come from. Sinha told Forbes that robotics has too often been treated as "a second-class citizen," with companies training on web-scale data and then fine-tuning on roughly "a hundred hours" of robot demonstrations. "That principle is so fundamentally broken," he said. "You need to see your most important tokens from step zero."

In practice, 1X wants Neo's training diet to include more than camera footage. Forbes describes a mix of live visual streams, proprioceptive data about joint position and forces, pressure and force data from the robot's hands, on-policy robot rollouts, simulation and human video. Sinha's shorthand to Forbes was: "Good tokens in, good tokens out."

That is a more expensive path than software-only model training because it depends on hardware in the world. It also explains why 1X is framing manufacturing as part of its AI strategy, not just operations.

Why Neo's body matters to 1X's model strategy

1X's specific claim is that Neo's human-like embodiment makes human video more useful as training data. "You build your embodiment so close, as small an embodiment gap as possible," Bornich told Forbes. "So now you can just pretrain all of the human video out there, and that actually transfers to your robot."

Forbes reports that Neo is tendon-driven rather than geared and that its hand has 22 actuated degrees of freedom. Bornich called the hand "the final boss of robotics," a phrase that captures both the technical problem and the commercial stakes: a household or general-purpose humanoid that cannot reliably grip, release and manipulate objects is not yet general-purpose.

Sinha also tied the hardware design to the data problem. The difference between gripping a bottle just firmly enough and not firmly enough is "tremendous," he told Forbes, and a camera alone cannot capture it.

The data flywheel is still mostly a claim

The strategic loop 1X is describing is familiar in AI, but harder in robotics: build more robots, deploy more robots, collect richer embodied data, train better models, make Neo more capable, then deploy more units. Forbes says 1X is scaling production and getting closer to shipping at scale, and says Neo is built with significant vertical integration at a Hayward, California facility, including from "literally raw copper wires," a description attributed to Sinha.

What is not in the record is just as important. 1X has not disclosed the World Model Lab's budget, size, hiring plan or location. The report also does not include Neo production volume, current deployment count, customer count, shipping timeline, revenue or valuation.

That leaves the central question open: whether 1X can turn its hardware ambitions into enough real-world robot data before one of the hundreds of humanoid robotics rivals gets there first. Forbes columnist John Koetsier says he is tracking more than 400 humanoid robotics companies, and the number matters because the next phase may be less about a single impressive demo than about who can afford the slow, physical work of gathering the right data at scale.