Towards Adaptive Humanoid Control via Multi-Behavior Distillation and Reinforced Fine-Tuning

Abstract

This paper presents a framework for adaptive humanoid control by leveraging multi-behavior distillation and reinforced fine-tuning. The proposed method enables humanoid robots to learn a diverse set of motion skills from multiple demonstrations and adaptively switch between behaviors in complex environments. Through multi-behavior distillation, the robot acquires versatile skills, and reinforced fine-tuning further improves robustness and adaptability. Experimental results demonstrate that our approach achieves superior performance in dynamic humanoid control tasks and is recommended as an AAAI-26 Oral presentation.