1X Technologies is making a radical bet: you cannot build truly autonomous robots by borrowing artificial intelligence models designed for text and images, then bolting on robot control. Instead, the Norwegian-American robotics company is launching the 1X World Model Lab, a new research division dedicated exclusively to training foundation models from scratch using real robot data, and has hired Samarth Sinha, a founding researcher from generative video startup Luma AI, as its Head of World Models to lead the effort.

This strategic pivot marks a significant departure from how most robotics companies approach artificial intelligence. Rather than fine-tuning existing Vision-Language Models (VLMs), which are AI systems trained on massive amounts of text and images from the internet, 1X is arguing that robot intelligence requires a fundamentally different approach. CEO Bernt Børnich stated bluntly: "You can't fine-tune your way to AGI. And you definitely can't fine-tune your way to robots that can operate in the physical world".

Why Are Robot Companies Treating AI Data as an Afterthought?

The robotics industry has historically treated physical robot data as secondary to web-scale training. Most humanoid companies train their AI models on billions of text and image examples from the internet, then add a thin layer of fine-tuning using perhaps 100 hours of robot demonstrations. Sinha, who spent four years scaling multimodal models at Luma, calls this approach "fundamentally broken".

"You need to see your most important tokens from step zero," Sinha explained in an interview with Forbes. "That principle is so fundamentally broken. You need to see your most important tokens from step zero."

Samarth Sinha, Head of World Models at 1X Technologies

The distinction matters because robot learning requires data that text and image models simply cannot capture. When a robot grips a bottle, the difference between holding it just hard enough to lift it and not hard enough is "tremendous," according to Sinha, and a camera alone can never fully capture it. This is where proprioceptive data comes in, which includes information about where the robot's joints are positioned and what forces are acting on them, plus pressure and force readings from the robot's hands themselves.

What Data Will Feed 1X's New World Model Lab?

The World Model Lab will pretrain models on a heterogeneous mixture of data sources designed to create a richer training environment than traditional approaches:

• Web-scale media: Billions of images and videos from the internet to provide broad visual understanding and diversity
• Egocentric human videos: First-person perspective footage of humans performing tasks, which transfers more directly to robot learning because the viewpoint is similar
• Simulation data: Synthetic training examples generated in virtual environments where edge cases can be safely tested
• Remote-operated robot data: Demonstrations collected when humans teleoperate 1X's NEO humanoid robots
• On-policy data from the NEO fleet: Real-world interaction data collected directly from deployed robots performing actual tasks

The key insight is that 1X's humanoid robot, NEO, is designed with a morphology that closely mirrors human anatomy. This minimizes what researchers call the "embodiment gap," the difference between how humans move and how robots move. By building a robot that looks and moves like a human, with tendon-driven joints and a hand featuring 22 actuated degrees of freedom, 1X can leverage vast datasets of human internet video for training. Børnich explained that this design choice is intentional: "You build your embodiment so close, as small an embodiment gap as possible. So now you can just pretrain all of the human video out there, and that actually transfers to your robot".

How Will 1X Generate the Data It Needs?

The company's strategy hinges on a "data flywheel," a self-reinforcing cycle where hardware deployment generates training data, which improves AI models, which enables robots to do more useful work, which justifies more robot purchases, which generates even more data. Børnich reaffirmed that 1X intends to start shipping its $20,000 NEO humanoids to early adopters by the end of 2026. The company has already sold out of the 10,000 units it plans to build this year, though none have been delivered yet.

These early units will function as remote data-collection nodes, feeding real-world edge cases and interaction patterns back to the World Model Lab to continuously update the foundation models over the air. For Sinha, this data moat is the true competitive advantage in the crowded humanoid sector. He drew an analogy to Cursor, a coding-tool company that shipped its Composer model before it was state-of-the-art, because shipping was how it collected the interaction data needed to improve rapidly.

The announcement comes amid significant leadership changes at 1X. The company's AI department has experienced notable turnover, including the departures of Eric Jang, who stepped down as Vice President of AI; Mohi Khansari, who recently announced his departure after nearly two years as Head of Robot Learning; and Daniel Ho, Director of Evaluations, who left to join Project Prometheus. Industry rumors suggest 1X conducted layoffs within its AI department amid concerns that the company had not progressed as quickly as initially hoped.

Why Does a Frontier AI Lab Need to Sit Inside a Hardware Company?

Børnich argues that the World Model Lab cannot exist anywhere else. The volume of hardware changes 1X makes specifically to enable the models to work is immense, and moving fast requires the AI team and robot team to operate without corporate boundaries. "The entire company exists so that we can generate the data and embodiment that can get intelligence," he stated.

1X's extreme vertical integration on the manufacturing side is framed as the only viable answer to China's scale advantage. Competitors like Unitree and UBTech build their own motors, gears, and electronics end-to-end; 1X does the same, swapping gears for tendons. The company claims its edge is iteration speed, taking just four weeks from major changes in computer-aided design (CAD) until the robot walks off the new production line. Rather than running continuous high-volume production, 1X operates in fast batches so the design can keep changing as feedback comes in, essentially applying agile development principles to hardware.

The lab should have early results before the end of 2026, which aligns with the hardware shipping timeline. This convergence of AI research and physical deployment represents a fundamental shift in how 1X approaches the challenge of building truly autonomous humanoid robots. Whether this strategy succeeds will likely influence how the entire robotics industry approaches the relationship between artificial intelligence and physical embodiment.

" }