Trajectory planning with reinforcement learning for a spatial under-constrained cable-driven parallel robot

In this thesis, trajectory planning is carried out using a reinforcement learning method, the Deep Q-Network (DQN), to ensure that the end-effector and cables of an under-constrained spatial cable-driven parallel robot move in the workspace without colliding with obstacles. For this purpose, reinforcement learning training is performed on the robot model created in the simulation environment, and the end-effector and cables reach the target without contacting obstacles using the reinforcement learning agent obtained as a result of the training. The artificial neural networks trained in the simulation environment are used on the developed experimental CDPR system, thus testing their performance in the real environment. The position and size information of real-world objects is obtained by processing the depth and color data from an RGB-D camera, and these data are provided to the reinforcement learning agent as observations. In these tests, the agent reaches the target without collision with a success rate of 98% in an obstacle-free environment and 80% in an environment with an obstacle. In addition, the agent trained for the obstacle environment is applied to a sample pick-and-place scenario, thereby demonstrating the applicability of the system in real tasks and its adaptability to actual operating conditions.