This study developed a robotic arm self-learning system based on virtual modeling and reinforcement learning. Using the model of a robotic arm, information concerning obstacles in the environment, initial coordinates of the robotic arm, and the target position, this system automatically generated a set of rotational angles to enable a robotic arm to be positioned such that it can avoid all obstacles and reach a target. The developed program was divided into three parts. The first part involves robotic arm simulation and collision detection; specifically, images of a six-axis robotic arm and obstacles were input to the Visualization ToolKit library to visualize the movements and surrounding environment of the robotic arm. Subsequently, an oriented bounding box algorithm was used to determine whether collisions had occurred. The second part concerned machine-learning–based route planning. The TensorFlow was used to establish a deep deterministic policy gradient model, and reinforcement learning was employed for the response to environmental variables. Different reward functions were designed for tests and discussions, and the program’s practicality was verified through actual machine operations. Finally, the application of reinforcement learning in route planning for a robotic arm was proved feasible by the experiment. Such an application facilitated automatic route planning and achieved an error of less than 10 mm from the target position.