Soft Computing towards Inverse Kinematics

• Gustavo H. G. Ledandeck UFABC
• Mateus Coelho Silva UFABC

Resumo

The Inverse Kinematics Problem (IKP) is a cornerstone of robot control, yet its complex, non-linear nature presents ongoing challenges. While deep learning has emerged as a powerful tool for approximating the IKP, the relationship between network architecture and performance is not always clear. This paper presents a systematic evaluation of feedforward neural network depth and architecture for solving the IKP for 3-DOF and 4-DOF robotic manipulators. Two Multilayer Perceptrons (MLPs) were compared with the TensorFlow (TF) model featuring Batch Normalization against classic Multilayer Perceptrons using Scikit-Learn (SKL) without Batch Normalization across a range of 3 to 10 hidden layers. Performance was evaluated on Euclidean distance error, accuracy under multiple tolerances, and joint-level prediction metrics (R2, RMSE, MAE). The results provide compelling empirical evidence that shallower architectures (3-4 hidden layers) consistently outperform deeper networks for this task, regardless of the specific architecture. Furthermore, the simpler SKL architecture demonstrated a statistically significant performance advantage over the TF, suggesting that for specific problem classes, additional regularization layers may not be beneficial. This work provides foundational guidelines for designing efficient and accurate neural network-based IKP solvers.

Palavras-chave: Accuracy, Systematics, 3-DOF, Neurons, Computer architecture, Kinematics, Multilayer perceptrons, Manipulators, Batch normalization, Guidelines, Keywords: Inverse Kinematics, Soft Computing, Neural Networks, Deep Learning, MLP, Robotics, Manipulator Control, Architecture Optimization, TensforFlow, Scikit-Learn