Leveraging Symmetries in Gaits for Reinforcement Learning: A Case Study on Quadrupedal Gaits

In this research, we address the complex task of developing versatile and agile quadrupedal gaits for robotic platforms, a domain predominantly governed by model-based trajectory optimization methods. We propose an innovative, reference-free reinforcement learning framework that exploits the intrinsic symmetries of dynamic systems to synthesize a broad array of naturalistic quadrupedal locomotion patterns. By capitalizing on distinct symmetry characteristics - namely temporal, morphological, and time-reversal - our approach efficiently facilitates the generation and transition among diverse gaits such as pronking, bounding half-bounding and galloping, across a spectrum of velocities, circumventing the necessity for expert-generated trajectories or complex reward structures. Implemented on the Petoi Bittle robotic model, our methodology illustrates robust and adaptable gait generation capabilities, significantly broadening the scope for robotic mobility and speed adaptability. This contribution not only advances our comprehension of quadrupedal locomotion mechanisms but also underscores the pivotal role of symmetry in the development of scalable and effective robotic gait strategies. Our findings hold substantial implications for robotic design and control, potentially enhancing operational versatility and efficiency across a variety of deployment environments.