Realizing "Human-Like" Control of Humanoid Robots through Imitation Learning Based on World Models

Abstract

In recent years, imitation learning — particularly a method called Generative Adversarial Imitation Learning (GAIL) — has been utilized to achieve human-like walking control. However, the model-free reinforcement learning used in this process suffers from low sample efficiency and limited robustness in complex environments. Since humans can adapt their walking even in complex environments, providing robots with a similar level of robustness is crucial for realizing human-like walking control. Therefore, this research aims to achieve human-like and robust walking control by integrating world models, known for their high sample efficiency and robustness, with imitation learning. In the experiment, fast simulations were conducted on a GPU using Isaac Gym. As a result, when the integrated algorithm attempted to learn human-like walking, it unexpectedly acquired behaviors such as "curling the entire body and jumping" or "fully extending the legs and standing rigidly." This paper reports these results and discusses potential causes.