Learning-Based Approaches for Forward Kinematic Modeling of Continuum Manipulators I. Mahamat Loutfi ∗ A.H. Bouyom Boutchouang ∗∗ A. Melingui ∗∗ O. Lakhal ∗∗∗ F. Biya Motto ∗ R. Merzouki ∗∗∗ ∗ Department of Physic’s, Faculty of Sciences, University of Yaounde I, Yaounde 8390, Cameroon (loutfimrane@gmail.com. ∗∗ Department of Electrical and Telecommunications Engineering, ENSP, University of Yaounde I, Yaounde 8390, Cameroon ( achillemelingui@gmail.com) ∗∗∗ CRIStAL Laboratory, CNRS-UMR, Villeneuve d’Ascq 59655, France Abstract: Forward kinematic model (FKM) is an essential module in the control law design of manipulator robots. Unlike rigid manipulators where it can be easily established, it remains a real challenge for their continuum counterparts. Model-based and learning-based approaches are commonly used for the forward kinematic modeling of continuum manipulators. Model- based approaches generally lead to imprecise FKM models due to several modeling assumptions, while learning-based approaches generally yield acceptable performance. However, the choice of an appropriate learning model remains a challenging task. In the framework of the forward kinematic modeling of continuum manipulators, this paper proposes an experimental and structural comparative study of the commonly used learning models, namely the multilayer perceptron (MLP), radial based functions (RBF), support vector regression (SVR), and Co-Active adaptive neuro-fuzzy inference system (CANFIS). The Compact Bionic Handling Assistant (CBHA) robot is used as an experimental platform and the predictions of the different learning models are compared respectively to a high precision motion capture system. According to the comparative study, we noted better accuracy for SVRs, rapid convergence for RBFs, and a good compromise between learning time and accuracy for MLPs. CANFIS offers accuracy close to that of SVRs but with much shorter learning time. Keywords: Multi-section Continuum Manipulators, kinematic modeling, Multilayer Perceptron (MLP), Radial Basis Function (RBF), Support Vector Regression (SVR), and Co-Active Adaptive Neuro-Fuzzy Inference System (CANFIS). 1. INTRODUCTION Machine learning is a tool that is increasingly used in the study of new generations of robots that combine mechanical flexibility, material elasticity and lightness. Such robots mimic the behavior of living beings as octopus arms, muscles, tentacles , elephant trunks, and cephalopod members. These properties make them uniquely suited for a large number of applications, including surgery, underwater operations, and exploration. Continuum manipulators are designed with flexible ma- terials and inherit non-stationary behaviors due to the hysteresis effects of some actuators, viscoelasticity and the loss of certain physical properties of the materials that they are made from. Unlike their rigid counterparts, which are made of rigid bodies, and from which the FKM can easily be derived, these characteristics make their kine- matics difficult to establish. Contributions on the FKM of continuum manipulators can be summarized into two main approaches, namely, model-based and learning-based approaches. Model-based approaches involve the estab- lishment of a kinematic model based on approximate as- sumptions about the physical structure of the manipulator robot. Regarding learning-based approaches, the paramet- ric space is divided into several groups according to the robot’s operating modes. A mathematical model derived from the learning algorithms makes it possible to establish a relationship between effects (expert observation, sensor measurements, and statistical data) and causes (input references). As regards to model-based approaches, various methods have been proposed to provide a solution to FKM of continuum manipulators. To establish the forward kine- matics of a continuum manipulator composed of 4 sec- tions, Hannan and Walker (2003) considered that the bending motion of a section of the manipulator could be better described using a constant curve (Webster III and Jones (2010)). Godage et al. (2011) used shape func- tions that incorporate the manipulator structure to model a multi-section variable-length continuum arm. Escande et al. (2012) used the constant curvature approach to model a two-section bionic manipulator mimicking the Preprints of the 21st IFAC World Congress (Virtual) Berlin, Germany, July 12-17, 2020 Copyright lies with the authors 10034