Neural-Network-Based Pose Estimation of Underactuated Cable-Driven Parallel Robots

Underactuated cable-driven parallel robots (UACDPRs) are a class of parallel manipulators in which traditional rigid links are replaced by fewer flexible cables than the end-effector degrees of freedom. Because of underactuation, conventional forward kinematics (FK) formulations are inherently underdetermined and necessitate the use of sensor redundancy to achieve a unique solution. Previous studies focused on sensor-fusion techniques to address the optimal use of redundant sensory data for pose estimation. This paper instead investigates how different combinations of sensory inputs affect the prediction accuracy of Feed-forward Neural Networks (FNNs). A real-world dataset acquired on a UACDPR in previous studies is divided into independent subsets used for training and validation, as well as comparison with the state-of-the-art results. This study demonstrates that combining cable lengths with IMU data yields the most balanced average 6D pose estimation, enabling optimally configured FNNs to outperform recent model-based iterative solvers in both accuracy and inference time.