Learning Quadrupedal Motion Through Multi-Objective Soft Actor-Critic

This paper presents a study on teaching a simulated quadruped with twelve joints to walk using the Soft Actor-Critic (SAC) algorithm. Furthermore, we introduce a multi-objective approach to efficiently integrate diverse and conflicting reward functions, enabling adaptive gait generation under different sets of preferences from a single training session. Leveraging SAC's adaptability and the efficiency of multi-objective integration, the presented approach is capable of generating locomotion without explicit model requirements while allowing for preference-oriented gait adaptations.