1X's new world model predicts the future for humanoid robots

Testing a humanoid robot in the real world is an absolute nightmare. Homes are messy, unpredictably arranged, and infinitely varied. If you had to physically verify every tweak to a robot's motor policy across thousands of cluttered living rooms, development would slow to a crawl.

Robotics company 1X unveiled the 1X World Model (1XWM), built to solve this exact bottleneck for its NEO humanoid platform.

Instead of staging physical dry runs or building painful, hand-crafted simulators, 1XWM enables NEO to anticipate the consequences of its physical actions inside an action-controllable neural simulation before making a move.

Controlled by joints, not just prompts

Unlike standard text-to-video tools like Sora that follow vague text prompts, 1XWM is steered by precise robot action trajectories. You feed it initial camera frames, sensory observations, and exact joint paths, and the model forecasts the latent representations of future frames along with the final success value of the task.

From a single starting frame, 1XWM can branch into distinct futures depending on what NEO actually does: grasping a mug, wiping down a counter, stepping backward, or strumming an imaginary air guitar.

Scaling up diverse physical data noticeably cleans up physical hallucinations. In early experiments with an air fryer, an undertrained 1XWM hallucinated the drawer and the chassis as a single fused block, showing the robot pulling the entire appliance off the counter. Once fed real interaction data, the model grasped that the basket slides out on a track and accurately simulated its mechanical boundaries.

1X is already turning this into a high-speed policy evaluation engine. According to 1X, even if two candidate policies have only a 15% gap in real-world success rates, a world model with 70% baseline accuracy can identify the superior policy with 90% confidence. It even allows counterfactual replay, testing alternative actions on states where a robot previously failed in the real world.

That said, the boundary is still obvious: feed it an unfamiliar, held-out object like a strange houseplant, and 1XWM still struggles to model the contact physics accurately.

Why it matters

We think 1X's bet here is extraordinarily pragmatic. Hand-coding physics engines for soft fabrics, shifting drawer contents, and greasy kitchen counters is notoriously brittle. Bypassing rigid physics engines in favor of learned video world models lets robot learning scale on GPU compute rather than manual asset design.

If synthetic sensory rollouts become indistinguishable from ground truth, the entire pace of humanoid development shifts. Robotics could experience its own synthetic data inflection point, similar to what frontier LLMs saw with reasoning and code.

However, hallucination remains the double-edged sword. If a world model misunderstands friction or gravity in a subtle edge case, policies evaluated inside it could pick up disastrous habits in silent confidence. Generalizing beyond a handful of kitchen counters into truly chaotic household environments will require astronomical amounts of real-world interaction data.