1X World Model

Humanoid company 1X's video world model, first used to evaluate policies, later used directly to control NEO.

The 1X World Model is a video-generation world model developed by the humanoid robotics company 1X. It was first announced in September 2024, trained on thousands of hours of video and action data collected by EVE robots in homes and offices: given the current image and a sequence of actions, it predicts the resulting video, and can simulate falling objects, cloth, doors, and drawers; 1X also released more than 100 hours of data and launched the 1X World Model Challenge. From June 2025 it was used to evaluate NEO's Redwood policy — running rollouts inside the model to compare policies and checkpoints while reducing real-robot testing. In January 2026, 1X connected it directly to NEO as a policy: a 14-billion-parameter text-conditioned video diffusion model first generates a video of the task being completed, and an inverse-dynamics model then infers actions from that footage; training used web video, 900 hours of first-person human video, and 70 hours of NEO data. Generating a 5-second video currently takes about 11 seconds, and dexterous tasks like pouring water remain difficult.

ExampleGiven the phrase “pull out a tissue” and the current image, 1XWM generates a video of NEO completing the action, and the inverse-dynamics model derives the actions for NEO to execute from it; generating several candidate videos in parallel and picking the best one raised this task's success rate from 30% to 45%.

Also called

1X World Model Challenge

Related

World Model · World-Model-based Policy Evaluation · Inverse Dynamics Model · Video Generation Model · 1X NEO · Redwood

Sources

1X World Model (1X, 2024-09)
1X World Model: evaluating Redwood AI (1X, 2025-06)
1X World Model | From Video to Action: A New Way Robots Learn (1X, 2026-01)

As of

2026-01