OMNIXTREME

北京通用人工智能研究院（BIGAI）、宇树、上海交通大学和中国科技大学等机构联合发布一项最新开源研究成果，该成果有望给人形机器人学习复杂运动的方式带来明显提效，且不必在动作保真度与可扩展性之间做艰难权衡。该研究提出了一种名为OMNIXTREME的新框架，成功让一个机器人学会执行包括后空翻、托马斯全旋、武术踢击在内的数十种高动态“极限运动”，并在宇树G1机器人上实现了真实世界的高成功率部署。OMNIXTREME研究团队提出了两阶段训练框架。实验结果表明，OMNIXTREME在包含LAFAN1和自建XtremeMotion极限运动库的综合测试中，追踪保真度远超现有基线方法。

A cutting-edge open-source research achievement has been co-released by institutions including Beijing Institute for General Artificial Intelligence (BIGAI), Unitree, Shanghai Jiao Tong University, and University of Science and Technology of China. This achievement is expected to significantly enhance the efficiency of humanoid robots in learning complex movements, eliminating the need for difficult trade-offs between motion fidelity and scalability. The research proposes a novel framework named OMNIXTREME, which enables robots to master dozens of highly dynamic "extreme sports" such as backflips, Thomas flair moves, and martial arts kicks, and has achieved high-success real-world deployment on the Unitree G1 robot. The OMNIXTREME research team presents a two-stage training framework. Experimental results show that OMNIXTREME exhibits considerably higher motion tracking fidelity than existing baseline methods in comprehensive benchmarks covering LAFAN1 and the self-built XtremeMotion extreme motion dataset.

搜集汇总

背景与挑战

背景概述

OMNIXTREME数据集专注于高动态人形机器人控制，通过统一策略实现多种极端行为的控制，包括后手翻、托马斯全旋等高难度动作。数据集采用预训练和后期训练相结合的方法，结合严格的电机约束和领域随机化，有效解决了模拟到现实的过渡问题。