Two-stage fuzzy object grasping controller for a humanoid robot with proximal policy optimization

Abstract

As science and technology have developed, an increasing amount of research on humanoid robots has been conducted. In this paper, a method based on deep reinforcement learning, optimization algorithms, and fuzzy logic for self-guided learning in humanoid robots is proposed. The method primarily relies on proximal policy optimization. The proposed model enables the humanoid robot to determine the optimal action on the basis of environmental feedback. A task was divided into two steps to train the optimal model for each step of the task; these models were then integrated. This division of the task was completed to prevent bias towards a single step. The performance of numerous optimization algorithms was evaluated, and the artificial bee colony algorithm was found to be the most successful algorithm for determining the optimal combination of parameters for the task. Deep reinforcement learning was demonstrated to be an effective method for enabling the humanoid robot to learn how to grasp objects and place them in target areas. The proposed learning method also combines optimization algorithms with fuzzy logic theory to further improve performance. The feasibility of the proposed method was validated through experiments.