[{"data":1,"prerenderedAt":96},["ShallowReactive",2],{"\u002Fen\u002Fglossary\u002Fworld-model":3},{"id":4,"title":5,"alternateName":6,"body":7,"description":86,"extension":87,"keywords":88,"meta":89,"navigation":90,"path":91,"seo":92,"stem":93,"updated":94,"__hash__":95},"glossary\u002Fglossary\u002Fen\u002Fworld-model.md","World Model","世界模型",{"type":8,"value":9,"toc":80},"minimark",[10,15,24,29,55,59,72],[11,12,14],"h1",{"id":13},"what-is-a-world-model","What Is a World Model?",[16,17,18,19,23],"p",{},"In robotics, a ",[20,21,22],"strong",{},"world model"," is an internal predictive model of environment dynamics learned by an agent: given the current observation (images, state) and an intended action, it predicts what the environment will look like next. With a world model, a robot can roll out the consequences of actions \"in its head\" — imagine first, act second — instead of trial-and-erroring every step in the real world.",[25,26,28],"h2",{"id":27},"relationship-to-vla-and-reinforcement-learning","Relationship to VLA and Reinforcement Learning",[30,31,32,44],"ul",{},[33,34,35,43],"li",{},[20,36,37,38],{},"Versus ",[39,40,42],"a",{"href":41},"\u002Fen\u002Fglossary\u002Fvla-model","VLA models",": a VLA maps observations and instructions directly to actions — a reactive policy. A world model explicitly captures \"action → consequence\" and supports planning and rollouts. The two are complementary: a world model gives a VLA the ability to imagine and verify.",[33,45,46,49,50,54],{},[20,47,48],{},"Versus reinforcement learning",": in model-based RL, the policy trains largely by \"dreaming\" inside the learned world model, requiring far fewer real interactions and cutting the cost of real-robot sampling. This parallels training in a simulator followed by ",[39,51,53],{"href":52},"\u002Fen\u002Fglossary\u002Fsim-to-real","Sim-to-Real"," transfer — a world model is essentially a differentiable \"neural simulator\" learned from data.",[25,56,58],{"id":57},"why-it-is-a-2026-embodied-ai-focus","Why It Is a 2026 Embodied-AI Focus",[16,60,61,62,66,67,71],{},"First, progress in video generation shows large models can absorb substantial physical regularities from massive video corpora — a foundation for general world models. Second, real-robot data is scarce and expensive, so the field wants world models to amplify limited ",[39,63,65],{"href":64},"\u002Fen\u002Fglossary\u002Fteleoperation","teleoperation"," data by generating and evaluating additional virtual experience. Third, ",[39,68,70],{"href":69},"\u002Fen\u002Fglossary\u002Fembodied-ai","embodied AI"," has high safety stakes: rehearsing inside an internal model before executing on hardware is a natural way to reduce risk.",[16,73,74,75,79],{},"Whatever the technical route, world models must ultimately be validated on real hardware. BXI's ",[39,76,78],{"href":77},"\u002Fen\u002Frobots\u002Fhumanoid-robot","humanoid robot"," ships with MuJoCo simulation and a ROS2 SDK, supporting the full loop of in-model training followed by real-robot validation.",{"title":81,"searchDepth":82,"depth":82,"links":83},"",2,[84,85],{"id":27,"depth":82,"text":28},{"id":57,"depth":82,"text":58},"A robotics world model predicts future environment states from the current state and an action, supporting planning, simulation, and policy training.","md","world model, embodied AI, predictive model, model-based reinforcement learning, robot learning",{},true,"\u002Fglossary\u002Fen\u002Fworld-model",{"title":5,"description":86},"glossary\u002Fen\u002Fworld-model",null,"M9XOHg398s6hJe7UWdj2CPKScagecDXq33iVYNAg-_U",1785156467527]