Parallel Kinematic Machines (PKMs) have been extensively studied in modern robotics due to their structural advantages, such as high stiffness, greater load capacity, and superior precision compared to serial manipulators. However, these advantages come with challenges, particularly regarding the complexity of modeling and dynamic control, due to the presence of multiple closed kinematic chains and strong coupling between links In this work, the development of a computational model is proposed to simulate a planar parallel manipulator of the 3RRR type, whose prototype was developed and built at EESC-USP (FAPESP 2018/21336-0). A relevant aspect of this study is the consideration of link flexibility, a feature that more accurately reflects the physical reality of the system but, on the other hand, makes analytical modeling even more challenging. Flexibility introduces nonlinear dynamic effects and elastic-structural couplings, requiring advanced numerical methods or alternative data-driven approaches for simulation and control. To address these difficulties, Artificial Neural Networks (ANNs), specifically Multi-Layer Perceptrons (MLPs), were adopted to build a predictive model of the manipulator. Eight experiments were conducted on the prototype, in which motor angles and strain signals obtained from strain gauges installed on the flexible links were recorded. These six signals were used as inputs to the MLP, while the end-effector pose data constituted the output targets. The MLP architecture was iteratively optimized by evaluating different combinations of the number of neurons, hidden layers, and learning rates. The training process incorporated early stopping techniques to prevent overfitting. The obtained results show that the MLP was able to successfully learn the relationship between input signals and the manipulator's pose, even in the presence of noisy data from the experimental environment. The neural network's predictions closely followed the actual trajectories, indicating the feasibility of using ANNs for modeling complex flexible manipulators. Future work includes exploring more sophisticated architectures such as recurrent or temporal convolutional networks, as well as integrating hybrid physics-informed models (Physics-Informed Neural Networks – PINNs), which can embed the system's physical equations to further enhance the model's generalization capability.
