Comparative Study of Learning and Execution of Bipedal by Using Forgetting Mechanism in Reinforcement Learning Algorithm

Abstract

A bipedal which resembles humans are programmed for performing specific tasks. The proposed work scope was to design, program, and validate RL based algorithms for navigation of Bipedal walking. The bipedal navigation implements forgetting mechanism in the traditional Q-learning algorithm which results in learning walk without prior learned dynamics of the system. Simulations were carried out for all three joints of each leg for evaluating the feasibility of the forgetting mechanism algorithm, the optimal policy, and the optimal actions for navigation. The reinforcement control algorithms for bipedal had been applied to take the self-decision. Bipedal senses the current state and moves to the goal state by learning or by using data stored in the lookup table in the execution phase. This reduces the learning and execution number of iterations of the bipedal by a considerable amount but total learning and execution time remains approximately the same. Simulation is done on the MATLAB platform and SimSpace Multibody dynamics toolbox to verify results. The bipedal model performs object identification, object localization in the dynamic environment, learns through the RL controller then executes to reach the object identified.