SkillBlender: Towards Versatile Humanoid Whole-Body Control via Skill Blending

Yuxuan Kuang, Amine Elhafsi, Haoran Geng, Marco Pavone, Yue Wang

Keywords: Humanoid, Whole-Body Control, Loco-Manipulation, Bimanual Manipulation, Reinforcement Learning

TL;DR: In this paper, we introduced SkillBlender, a pretrain-then-blend framework for versatile and robust humanoid whole-body control.

Abstract: Humanoid robots hold significant potential in assisting humans across diverse environments and tasks thanks to their flexibility and human-like morphology. However, whole-body control remains a significant challenge, given the high-dimensional action space and the inherent instability of bipedal systems. Previous works often rely on either precise dynamic models with computationally expensive optimization or task-specific training with extensive reward tuning. In this work, we introduce **SkillBlender**, a hierarchical reinforcement learning framework that first develops a set of primitive skills using pre-designed dense rewards, and then reuses and blends these skills to accomplish more complex new tasks, requiring minimal task-specific reward engineering. Our simulated experiments on two complex loco-manipulation tasks show that our method significantly outperforms all baselines, while naturally regularizing behaviors to avoid reward hacking, resulting in more feasible and human-like movements.

Submission Number: 21