IJCAS

Unlike conventional rigid-link parallel robots, cable-driven parallel robots (CDPRs) have distinct advantages, including lower inertia, higher payload-to-weight ratio, cost-efficiency, and larger workspaces. However, because of the complexity of the cable configuration and redundant actuation, model-based forward kinematics and motion control necessitate high effort and computation. This study overcomes these challenges by introducing deep reinforcement learning (DRL) into the cable robot and achieves compensated motion control by estimating the actual position of the end-effector. We used a random behavior strategy on a CDPR to explore the environment, collect data, and train neural networks. We then apply the trained network to the CDPR and verify its efficacy. We also addressed the problem of asynchronous state observation and action execution by delaying the action execution time in one cycle and adding this action to be executed to match the motion control command. Finally, we implemented the proposed control method to a high payload cable robot system and verified the feasibility through simulations and experiments. The results demonstrate that the end-effector position estimation accuracy can be improved compared with the numerical model-based forward kinematics solution and the position control error can be reduced compared with the conventional open-loop control and the open-loop control with tension distribution form.

Keywords: Cable-driven parallel robot, deep reinforcement learning, motion control.