Unleashing Humanoid Reaching Potential Via Real-World-Ready Skill Space
Zhikai Zhang, Chao Chen, Han Xue, Jilong Wang, Sikai Liang, Yun Liu, Zongzhang Zhang, He Wang, Li Yi
AI summary
Problem
Learning complex whole-body control for large-space reaching from scratch is difficult due to optimization challenges and poor sim-to-real transferability, especially when coordinating multiple skills.
Approach
The authors propose Real-world-Ready Skill Space (R2S2), which constructs a library of real-world-ready primitive skills and uses heterogeneous skill ensembling (combining imitation and reinforcement learning in a CVAE framework) to create a unified neural skill representation for seamless coordination and planning.
Key results
- Autonomous whole-body reaching on Unitree G1 and H1 humanoids in simulation and real world
- Sim2real transferability with minimal reward engineering
- Unified continuous latent skill space for efficient high-level planning
- Large-reach-space whole-body teleoperation support
Why it matters
Provides a scalable, transferable framework for humanoid whole-body control that bridges the gap between simulation and reality, accelerating real-world deployment for daily assistance tasks.
Abstract
Humans possess a large reachable space in the 3D world, enabling interactions with objects at varying heights and distances. However, realizing such large-space reaching on humanoids is a complex whole-body control (WBC) problem. Learning from scratch often leads to optimization difficulty and poor sim2real transferability. To address these challenges, we present Real-world-Ready Skill Space (R2S2), a structural skill Manuscript received: July, 19, 2025; Revised October, 23, 2025; Accepted November, 25, 2025. This paper was recommended for publication by Editor Abderrahmane Kheddar upon evaluation of the Associate Editor and Reviewers’ comments. *Zhikai Zhang, Chao Chen, and Han Xue are co-first authors. 1First Author, Third Author, Sixth Author, and Ninth Author are with IIIS, Tsinghua University, China. 2Fourth Author and Eighth Author are with Peking University, China. 3First Author, Second Author, Third Author, Fourth Author, Fifth Author, Sixth Author, and Eighth Author are with Galbot, China. 4Ninth Author is with Shanghai AI Laboratory, China. 5Ninth Author is with Shanghai Qi Zhi Institute, China. 6Seventh Author is with Nanjing University, China. 7Fifth Author is with Tongji University, China. Digital Object Identifier (DOI): see top of this page. prior that helps autonomous whole-body-control task execution in an efficient manner while maintaining sim2real transferability. Inheriting knowledge from a set of real-world-ready primitive skills to ease multi-skill learning, R2S2 further expands the capability of primitive skills and learns a unified structural skill representation. By sampling from R2S2, we unleash humanoid reaching potential in many real-world tasks. As a beneficial side effect, R2S2 can also support humanoid whole-body teleoperation with a large reachable space. We validate the generalizability of R2S2 in various challenging goal-reaching tasks across different robot platforms, simulation and real world. We show some ex- amples in Figure 1. Project page: https://zzk273.github.io/R2S2/.