- Published on
1X Reveals Its 'World Model,' A Digital Twin to Accelerate Humanoid AI Training
- Authors
- Name
- Humanoids daily
- @humanoidsdaily
After a week of announcements detailing its Redwood AI brain and the underlying mobility controller for its NEO humanoid, robotics firm 1X has unveiled the next piece of its development puzzle: a "World Model" designed to dramatically speed up how its robots learn.
In a blog post and video, the OpenAI-backed company revealed the 1X World Model (1XWM), a generative AI system that functions as a data-driven digital twin of reality. Its purpose is to solve one of the most significant bottlenecks in robotics: how to efficiently test and evaluate an AI model's performance. Instead of running a robot through thousands of physical scenarios—a process that could take "several lifetimes"—1X can now simulate the future in what it calls "a bridge between the world of atoms and the world of bits."
More Than a Video Generator: An "Action-Controllable" Simulator
At first glance, the 1XWM looks like a video generation model. It creates short video clips of the NEO robot performing tasks. However, unlike text-to-video models such as Sora that respond to linguistic prompts, the 1XWM is action-controllable.
This means it's steered by the robot's exact, low-level action trajectories. A developer can provide the model with an initial scene (a few real video frames) and then feed it multiple different action sequences. The model then predicts and generates distinct video futures for each sequence—showing what would happen if NEO tried to grab a mug versus wipe a counter, all from the same starting point.
This precise control is what allows the world model to function as a powerful evaluation tool.
1X World Model Scaling Evaluation for Robots
Solving the "Evaluation Problem"
The primary application for the 1XWM is to accelerate iteration. As stated in the announcement, the system allows 1X to "quickly iterate on architectural decisions" and "select the best checkpoint from training runs."
For example, to determine whether including proprioception (the robot's own sense of joint position and movement) improves a policy, the team no longer needs to run extensive real-world tests. They can run both policies—one with proprioception and one without—through the world model across a curated dataset of scenarios.
Crucially, 1X reports a high correlation between the success rates predicted by the world model and the actual task scores from real-world evaluations. "Given a true real-world success rate gap of 15% between two policies, a World Model with 70% accuracy can accurately predict the better policy with 90% success," the company states. This allows their development feedback loop to shrink from weeks to minutes.
"We see enough signal that these evaluations are well correlated," noted Daniel Ho of the 1X AI team, confirming the model's practical use in making production-level decisions.
The Right Data is Key
The effectiveness of the 1XWM hinges on the data it's trained on. While the company uses various data sources, the team found that the most valuable input for improving the model's accuracy and understanding of physics is autonomous robot rollouts.
This data, captured as NEO robots attempt tasks in real-world environments like offices and homes, is particularly potent because it includes failures. "Failure data is really important for improving the model because it teaches the AI what not to do," a narrator explains in the company's video.
As the model is trained on more diverse interaction data, its grasp of physics becomes more nuanced. An early version of the model might see an air fryer as a single blocky unit. After training on data of a robot opening and closing it, the model learns that the tray is a separate, sliding component.
Still, the company is transparent about the model's limitations. It currently struggles to accurately simulate interactions with "held-out" objects that it hasn't seen in its training data.
The End Game: A Unified Data and Evaluation Engine
The 1XWM is more than just a testing tool; it points toward a more ambitious future. Eric Jang, VP of AI at 1X, posed a forward-looking question: "Consider the implications of what happens when the data generated by 1XWM... become indistinguishable from the real data."
If the synthetic data becomes as good as real data, the world model could be used not just for evaluation but also for generating vast amounts of high-quality training data. This would create a powerful, self-improving loop—a single, unified model for both "robot data + robot evals," as Jang put it.
This strategy aligns with the company's broader mission, as articulated by CEO Bernt Børnich, to find the "shortest path" to artificial general labor. That path, for 1X, appears to be paved with massive amounts of diverse, real-world data, collected by physically safe robots and amplified by increasingly sophisticated simulations like the 1X World Model.