Researchers at the Beijing Innovation Center of Humanoid Robotics, known as X-Humanoid, described Pelican-Sim 1.0 in a technical report posted on Sept. 10, 2026. It is a world model, an AI that learns how the world behaves so it can predict what happens next. Given a robot's camera view and a planned series of moves, it predicts the images the camera would show next. Robot labs still lean on scarce sessions in which people steer robots remotely, and on slow physics simulations. A simulator that learns from video could cut how much real robot time each new skill needs, if the gains hold up outside the team's own tests.
Pelican-Sim 1.0 is not a chatbot, and it does not control a robot by itself. Other software can use the video it predicts to create training data, to judge how well a robot's control program is likely to do, to choose between possible moves, or to improve the program, without running every trial on real hardware.
The team says four design choices make this work. The model describes the moves of different robots in one shared format, covering arm joints, grippers and hands on both sides. It also takes in a rendered video of the robot's planned motion, lined up with the camera's view. It switches on only part of itself at a time, which the team says gives it room for many different kinds of motion. And a faster version, trained to imitate the slower original, makes each prediction in four steps instead of 35. The researchers say that makes it 5.67 times faster.
The team says it trained the model on about one million real and simulated robot runs, about 8,000 hours of video, from seven collections including AgiBotWorld Beta, RoboMIND and RoboTwin. On video the model had not seen before, the researchers say its predictions matched the real footage more closely than those of the strongest rival systems they retrained for comparison, in all three collections they tested. In tests on RoboTwin, adding 500 practice runs made by Pelican-Sim to 50 demonstrations for each task raised a robot control program's success rate from 70% to 93%, they say. They also report that its scores for five versions of a control program closely tracked how those versions actually did, and that it helped in picking a robot's next move and in improving its control program.
The team says it will release the model and the code to run it. As of Sept. 21, 2026, its public code page held only a readme and a license. Until the model is out, it is not clear whether others can match the gains, whether the comparison with rival systems was fair, or how well the model handles scenes unlike those in its tests.