1X Technologies announced on June 4, 2026 the launch of the 1X World Model Lab — a new research division focused on pretraining robotics foundation models from scratch. Samarth Sinha, formerly a founding researcher at Luma AI, has been hired to lead the lab. The move marks a clear departure from a strategy based on fine-tuning language models for robotics.
Key takeaways
- 1X World Model Lab is a dedicated research unit for pretraining robotics foundation models
- Samarth Sinha (Luma AI) hired as Head of World Models
- Company lost three senior AI leaders: Eric Jang, Mohi Khansari, and Daniel Ho
- CEO declares: fine-tuning language models for robotics is fundamentally broken
- Plans to ship thousands of NEO robots at 20,000 USD each in 2026 as a data flywheel for new models
Structural shift after leadership departures
1X Technologies announced the lab against a backdrop of significant turnover. Within a few months the company lost three senior AI leaders: Eric Jang (VP of AI), Mohi Khansari (Head of Robot Learning), and Daniel Ho (Director of Evaluations). Reports of layoffs in the AI department also circulated in the industry.
Hiring Sinha to head the new lab is more than filling a vacancy. He specializes in generative video models and pretraining on multimodal data — skills that directly translate into training models that understand the physical world as a continuous stream rather than a sequence of text tokens.
Fine-tuning as a dead end
CEO Bernt Bornich was direct in his declaration on X: you cannot reach AGI through fine-tuning, let alone through fine-tuning for robots operating in the physical world. This is a clear statement against the dominant VLA (Vision-Language-Action model) approach, which involves grafting action heads onto ready-made language or multimodal models.
You cannot fine-tune your way to AGI. And you definitely cannot fine-tune your way to robots that can operate in the physical world.
Bernt Bornich, CEO 1X Technologies
Sinha elaborates: the robotics industry treats physical data as a second-class resource, bolted on after training built around text and internet images. In his view this is a foundational mistake — models must learn on physical data from step zero, not as a post-hoc correction.
NEO as the data flywheel
The centerpiece of the strategy is a fleet of NEO robots shipped to homes. 1X has confirmed plans to deliver thousands of NEO units at 20,000 dollars each in 2026. The robot has a human-like morphology — soft body, tendon-driven joints, hands with 22 degrees of freedom. The company views this minimization of the embodiment gap as deliberate: data collected by a robot with proportions close to a human should transfer more easily to models pretrained on egocentric human video.
Each deployed unit is a potential node for collecting real-world edge-case data. If 1X meets its 2026 shipping targets, it would have one of the largest datasets of human-environment interactions among all humanoid companies.
Why it matters
The 1X restructuring exposes a divide: should foundation models for robotics grow out of LLMs and multimodal models pretrained on internet data, or must they be built from scratch on physical data? 1X, together with Generalist AI and Physical Intelligence, is forming an increasingly clear camp in favour of the second approach. On the other side stands Google DeepMind with Gemini Robotics, OpenAI with its nascent robotics division, and numerous academic VLA groups. The outcome over the next two to three years will determine whether physical AI is built on the shoulders of existing language models or requires its own infrastructure and training data.
What is next
- 1X has committed to delivering thousands of NEO robots to homes by end of 2026 per CEO statement to Forbes
- Sinha is tasked with building a pretraining program from scratch on heterogeneous data: web video, egocentric human video, simulation, teleoperation, and on-policy NEO fleet data
- No next model timeline announced — priorities are now recruitment and infrastructure
