1X's new world model teaches humanoid bots to make toast

Training a humanoid robot to do chores has usually meant putting a human in a VR headset to puppet every single movement dozens of times. 1X wants its bots to just watch videos and figure things out on their own.

Robotics startup 1X unveiled the 1X World Model, a physics-based AI model designed to help its Neo humanoid robot learn real-world dynamics. By combining web videos with text prompts, the model allows Neo to attempt tasks it was never explicitly trained to do.

Learning physics by watching clips

According to 1X founder and CEO Bernt Børnich, the company spent years building the world model alongside Neo's hardware. The goal was to bridge the gap between digital content and physical action.

Børnich claimed Neo can now take knowledge from internet-scale video and apply it straight to physical environments. In his view, the bot can translate prompts into brand-new movements without prior human demonstrations, marking the start of self-guided robot learning.

Beyond direct control, the world model gives 1X engineers visibility into how Neo imagines executing an action. That internal reasoning loop provides valuable synthetic data to refine future iterations of the software.

From big claims to air fryers

Claiming a bot can turn any prompt into action is obviously marketing stretch. As TechCrunch pointed out, you cannot tell Neo to drive a car and expect it to parallel park.

A company spokesperson later clarified the nuances: the world model allows Neo to attempt anything, but what it has actually mastered remains very modest. Right now, Neo can handle small kitchen chores like sliding an air fryer basket in and out, dropping bread into a toaster, and giving people high fives.

1X considers these micro-skills the foundational stepping stones toward executing more complex, multi-step domestic chores. The update comes at a high-stakes moment for the startup, which opened consumer preorders for Neo last October and plans to ship units this year. The company declined to share preorder numbers or a firm shipping date, beyond saying orders beat expectations.

Why it matters

For prospective buyers, this gives a clearer picture of what early home robots will actually look like. If an assistant requires manual calibration for every mug and drawer in your house, it will never scale. Self-learning through video could eventually make household setup effortless. But realistically, popping toast into a slot is a long way from folding laundry or cooking dinner. Anyone receiving a Neo this year should still expect a research platform rather than a reliable butler.

For the robotics industry, 1X is betting on a major shift in data economics. Teleoperating humanoids to collect training data is brutally slow and expensive. Leveraging the endless sea of internet video as a training sandbox could drastically cut development time, putting pressure on competitors like Tesla and Figure to prove their own foundation models in physical space.

We think 1X's strategy is undeniably clever. Tapping passive web video bypasses the hardest bottleneck in robotics today, which is physical data collection. However, seeing physical dynamics is very different from managing contact forces and friction in messy home environments. Allowing a robot to attempt anything also means it can fail in unpredictable ways, and 1X still has to prove these visual shortcuts translate to safe, reliable hardware control.