Optimal Gait Control of Soft Quadruped Robot by Model-based Reinforcement Learning

Abstract: Quadruped robots offer distinct advantages in navigating challenging terrains due to their flexible and shock-absorbing characteristics. This flexibility allows them to adapt to uneven surfaces, enhancing their maneuverability. In contrast, rigid robots excel in tasks that require speed and precision but are limited in their ability to navigate complex terrains due to their restricted motion range. Another category of robots, known as soft robots, has gained attention for their unique attributes. Soft robots are characterized by their lightweight and cost-effective design, making them appealing for various applications. Recent advancements have made significant strides in practical control strategies for soft quadruped robots, particularly in diverse and unpredictable environments. An emerging approach in enhancing the autonomy of robots is through reinforcement learning. While this approach shows promise in enabling robots to learn and adapt to their surroundings, it necessitates rigorous training and must exhibit robustness in real-world scenarios. Moreover, a significant hurdle lies in bridging the gap between simulations and reality, as models trained in idealized virtual environments often struggle to perform as expected when deployed in the physical world. This thesis aims to address these challenges by optimizing the control of soft quadruped robots using a model-based reinforcement learning approach. The primary goal is to refine the gait control of these robots, taking into account the complexities encountered in real-world environments. The report covers the implementation of model-based reinforcement learning, including simulation setup, reward design, and policy refinement. Results show improved training efficiency and autonomous behavior, confirming the method’s effectiveness in enhancing soft quadruped robot capabilities.It is important to note that this report provides a concise summary of the thesis results due to the word limit imposed by the Department of Machine Design. For a comprehensive understanding of the research and its implications, the complete version is attached separately here.