具身世界模型开源 让机器人学会“预演”未来

Core Insights - The rapid evolution of robotic movement capabilities is highlighted, with robots now able to perform complex actions like backflips and running, but understanding physical interactions remains a challenge [1] - The introduction of the WoW (World of Wonder) embodied world model allows robots to develop better imagination and execution capabilities, akin to human understanding [2] Group 1: WoW Embodied World Model - The WoW embodied world model enables robots to predict and understand physical interactions, such as anticipating the consequences of knocking over a cup, thereby connecting imagination with real-world execution [4] - Developed by a collaboration between the Beijing Humanoid Robot Innovation Center, Peking University, and the Hong Kong University of Science and Technology, the model is open to global researchers and developers [6] - The model can adapt to various robotic forms and scenarios, including home, retail, industrial, and logistics environments, and can simulate extreme situations for data collection [6] Group 2: Autonomous Evolution and Learning - The WoW model features a self-evolving capability, allowing robots to learn and improve through a virtual world that mimics real-world logic [7] - It employs a dual-model system combining the embodied world model for physical predictions and a visual language model for multi-modal understanding and task planning, creating a feedback loop for continuous learning [7] - A comprehensive benchmark for the embodied world model has been established, assessing core capabilities such as perception, prediction, reasoning, and decision-making [9]