Robo-Net: A Novel Reinforced Walking Biped Design Using an Augmented Random Search Approach

Model-Based reinforcement learning has always presented significant approaches to better strategies for system dynamics. This paper highlights a random search approach that eliminates the need for enormous data and demonstrates exceptional sampling efficiency compared to model-based strategies for training policies for bipedal walking robots. Specifically, this paper stresses the utilization of a simple model-free random search technique called Augmented Random Search (ARS), which trains policies for MuJoCo (Multi-Joint dynamics with Contact) tasks and exhibits extensive variations when hyperparameters and random seeds are adjusted. Unlike model-based methods, which require substantial computational resources and large datasets, ARS offers a simplified approach with reduced data requirements, making it more practical for real-world implementations. For instance, ARS requires 30% fewer data samples than Trust Region Policy Optimization (TRPO) and Proximal Policy Optimization (PPO) while achieving similar levels of task success. ARS addresses challenges such as handling noisy and incomplete data, optimizing rehabilitation strategies for exoskeleton robots, and ensuring robust performance in variable environments. Through simulations, the performance of the bipedal robot is analyzed in a deliberately designed environment. The results exhibit enhanced sample efficiency by optimally tuning several hyper-parameters of the algorithm and demonstrating robust adaptability to alternative environmental changes. As the characteristics of bipedal walking robots can be associated with human locomotion, various biomedical applications, as well as robotic deployment in extreme terrains and hostile environments, become feasible. This research provides an elementary step toward developing bipedal walking robots by enhancing control kinematics while maintaining algorithmic simplicity.