1X has released a new update for their NEO humanoid, shifting from traditional hard-coded simulators to a data-driven "World Model." It uses internet-scale video data fine-tuned on robot interactions to allow the robot to "hallucinate" the outcome of its actions before performing them. This approach aims to solve the evaluation problem in robotics: how to test a general-purpose AI across thousands of novel tasks without needing a manual digital twin for every scenario.
You can read more about it on their website: https://www.1x.tech/discover/world-model-self-learning