Unleashing Humanoid Reaching Potential via Real-world-Ready Skill Space
Unleashing Humanoid Reaching Potential via Real-world-Ready Skill Space
RA-L, 2025
Zhikai Zhang1,3*, Chao Chen3,6*, Han Xue1,3*, Jilong Wang2,3, Sikai Liang3,7, Yun Liu1,3, Zongzhang Zhang6, He Wang2,3, Li Yi1,4,5
1Tsinghua University, 2Peking University, 3Galbot, 4Shanghai AI Laboratory, 5Shanghai Qi Zhi Institute, 6Nanjing University, 7Tongji University
R2S2 Overview. We present Real-world-Ready Skill Space (R2S2), a skill space that encompasses and encodes various real-world-ready motor skills. 1) We decompose the complex WBC motor skills into a library of primitive skills, each separately tuned and sim2real evaluated. 2) We ensemble multiple primitive skills into a student policy with a variational information bottleneck. 3) We train high-level planning policies to sample from R2S2 to efficiently and stably accomplish real-world goal-reaching tasks.