光象科技联合清华发布世界模型 ActEffect，训练后退出部署链路
让世界模型只在训练中教机器人“想后果”，执行时不增加算力，仿真成绩亮眼，正迈向工业场景。
机器人训练时，一个受控世界模型会预测动作后果并优化策略，但任务执行时它便退场，无需展开未来搜索。光象科技与清华李升波课题组发布物理原生世界模型 ActEffect，在 LIBERO 取得 98.8% 成功率，并显著提升分布偏移下的表现。
正文摘录
Title: The world model steps aside after training — and robots become more capable Source: QbitAI Spending significant time training alongside a robot before it starts its job, only for it to be removed from the deployment pipeline the moment the robot actually begins working? In the past, a world model for a robot was like a portable "mental sandbox" — the robot would first use it to predict what might happen in the future, then decide how to act. The more the model "thought," the longer the inference chain typically became. But what if this model only exists within the training ground? Its sole job is to "check" the consequences of actions proposed by the robot itself, then feed those checks back as signals for policy optimization. Once training is complete, this controlled world model e…