Improving grasping actions in humanoid robots with deep reinforcement learning algorithms

One of the most basic skills that humanoid robots need to acquire in order to help people or replace them in environments such as homes, hospitals, and workplaces is the ability to grasp objects. In addition, since they interact with people, they should be able to explain the actions they perform. The aim of this thesis is to ensure that grasping, one of the human characteristics, is performed by humanoid robots. In line with this purpose, Deep Reinforcement Learning (DRL) algorithms were used to train the Robotis-OP2 humanoid robot. The robot's grasping actions were performed using the location of the joints, the force information received from the joints, and the images captured by the robot's own camera. The robot was enabled to perform grasping autonomously. First, the detection of the grasping point was calculated using advanced kinematic computations to verify that the robot achieved an appropriate grip. Then, DRL methods were used to ensure that the robot performed grasping autonomously. Deep Q Network (DQN) and Duello DQN were used from the DRL algorithms, and these methods were tested with different reward functions. For all methods used to enable the robot to learn from experience, the Deep Q Learning from Demonstrations (DQfD) algorithm, one of the imitation learning methods, was utilized. To explain the robot's actions, the Randomized Input Sampling for Explanation (RISE) algorithm, one of the Explainable Artificial Intelligence (XAI) methods, was applied. At the end of the study, it was observed that the robot successfully achieved the desired goals. It was found that the Duello DQN algorithm provided better results than the DQN algorithm. Additionally, it was concluded that the training process duration was reduced by using the imitation learning method. Through the use of XAI, explanations for the reasons behind the robot's actions were obtained, and it was confirmed that a safer grasping process could be achieved.