A Generalist Agent Learning Architecture for Versatile Quadruped Locomotion

Quadrupeds can generate various motor behaviors with the muscle synergies activated by the central nervous system. However, versatile locomotion for quadruped robots remains challenging due to the complexity of the high-dimensional limb dynamics with many physical constraints. Current approaches typically apply a dedicated policy or controller for each motor behavior, which requires the optimization of a large number of parameters and training process is complicated. In this paper, we propose a Generalist Agent Learning Architecture (GALA) to learn diverse motor behaviors simultaneously with a single policy network for quadruped locomotion. GALA significantly decreases the number of trainable parameters while producing appropriate motor behaviors by simply reactivating the generalist policy based on different sensory feedback and commands at run time. We experimentally analyze and demonstrate the versatile locomotion delivered by GALA on both simulated and real quadruped robots in various environments. The code is available at https://github.com/vsislab/GALA. Note to Practitioners—This paper presents a generalist agent learning architecture (GALA) in which a single policy network learns various motor behaviors simultaneously for quadruped robots. GALA technically integrates multiple critics actor learning with a generalist policy to generate a diversity of motor behaviors in accordance with high-dimensional sensory feedback and commands. Our method is easy to train with a much smaller number of trainable parameters for versatile locomotion. The policy can be explored and exploited in simulated environments and directly deployed on the real robot through domain randomization technique that has been applied to many robots for sim-to-real transfer. The GALA-based quadruped robot flexibly generates versatile motor behaviors to smoothly traverse a series of real-world environments in runtime.