Automated Calibration of Planar Cable-Driven Parallel Manipulators by Reinforcement Learning in Joint-Space

Benefiting from modularity, cable-driven parallel robots (CDPRs) are capable of being reconfigurable by changes in their attachment points and, therefore, significant changes in their kinematic structures. Due to their wide-range motion, measuring CDPRs' fixed attachment points location can be limiting. This paper tackles the problem of identifying the manipulators' geometry based on their interoceptive sensors by reinforcement learning. We propose using Jacobian matrix elements to map rewards and actions into joint space without the appearance of local minimums and multiple solutions of forward kinematics. Feasibility of this method is demonstrated by a planar redundant CDPR. Without an expensive tracking system, the robot is capable of autocalibration based on the cable length measurements (actuator feedback) and quantization factors of any configuration space, while keeping all the cables under tension force. For instance, if the workspace is discretized with a grid resolution of 1 cm, this algorithm is capable of reducing the initial error of 2.2 m, as low as 1 cm. Further extension of this method toward higher technology readiness levels can improve the possibility of commercializing these manipulators toward plug-and-play setups.