| dc.description.abstract | Control of robotics for object grasping and manipulation is still a complex problem with many different approaches and solutions. This project examines the usage of actor-critic reinforcement learning methods in an attempt to teach reinforcement learning agents to control a robotic manipulator arm attached to a mobile research robot. The agent controlling the arm is trained to attempt to reach for and assume a pre-grasp position around an object placed on a table in a simulated world. Different agents are trained with various degrees of utilization of the robot's sensory systems and with varying definitions of the parameters associated with reinforcement learning, such as the state space, action space, and reward function definitions. Comparative experiments are conducted in the same simulated environment, comparing the different reinforcement learning agents with each other, as well as comparing the best-performing agent with a traditional motion planning algorithm. The results indicate that the best-performing agent's proficiency at reaching and assuming a pre-grasp position in its initial simulation environment (PyBullet) peaks at 92%. However, when transferred to and retrained in the same environment in which the motion planning algorithm is implemented (Gazebo), its accuracy reduces significantly to 55.8%. In contrast, the motion planning algorithm achieves a success rate of 76% for the same task. Although the results suggest that the best-performing trained agent's accuracy is worse than that of the traditional motion planning algorithm for the task presented in the Gazebo environment, it is significantly faster at reaching the object because it does not require pre-planning the motion. Additionally, different simulation environments are discussed, highlighting the differences between using Gazebo, the most common simulation environment for robotics development, and PyBullet a physics engine implemented for ease of use in Python. | eng |