Core Viewpoint - The article discusses the release of a new open-source model named UnifoLM-WMA-0, which is designed to enhance the interaction between robots and their environments through a world model that understands physical laws [1][9]. Group 1: Model Performance - The model demonstrates effective performance in tasks such as stacking blocks, with predictions closely matching actual operations [3]. - It can also handle more intricate tasks, such as organizing stationery, showcasing its versatility [7]. Group 2: Model Features - UnifoLM-WMA-0 is part of the UnifoLM series, specifically tailored for general robot learning and adaptable to various robotic platforms [9]. - The model's training code, inference code, and checkpoints have been fully open-sourced, quickly gaining over 100 stars on GitHub [11]. Group 3: Training Strategy - The training strategy involved fine-tuning a video generation model using the Open-X dataset to adapt its capabilities to real-world robotic tasks [15]. - The model operates under a dual-function architecture: a decision mode for predicting key information during physical interactions and a simulation mode for generating realistic environmental feedback based on robot actions [20]. Group 4: Dataset Utilization - The training utilized five open-source datasets provided by Unitree Technology, which contributed to the comprehensive training process [22]. - The model excels as a simulation engine, capable of generating controlled interactions based on current scene images and future action commands [23].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 编辑丨具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要 的。 UnifoLM-WMA-0是宇树科技推出的开源世界模型-行动架构，该架构跨越多种机器人实体形态，专 为通用机器人学习而设计。其核心组件是具备理解机器人与环境间物理交互能力的世界模型，该模 型提供两大关键功能：（a）仿真引擎——作为交互式模拟器运行，为机器人学习生成合成数据； （b）策略增强——与行动模块连接，通过预测与世界模型的未来交互过程，进一步优化决策性能。 项目链接：https://unigen-x.github.io/unifolm-world-model-action.github.io/ 架构说明 UnifoLM-WMA-0是一种嵌入世界模型的策略架构。该框架使世界模型能够以两种模式运行：（1） 决策模式-预测未来物理交互信息以辅助策略生成动作；（2）模拟模式-根据机器人动作生成高保真 度的环境反馈。 对视频生成模型进行微调：首先，我 ...