Development of the Inverse Kinematics Model of Arslan Humanoid Robot Using Stereo Vision and CNN

This paper presents the design and implementation of an inverse kinematics (IK) model for Arslan, a humanoid robot with a biomechanically realistic neck structure. The neck, modeled from 3D-scanned human cervical vertebrae, achieves natural flexion/extension and axial rotation through a two-motor system coupled with springs and elastic bands. The resulting mechanical compliance introduces non-linearities that make conventional analytical IK models ineffective. To address this, a stereo vision system with a laser reference was used to capture over 2,500 head pose samples across defined motor angle ranges. A Convolutional Neural Network (CNN) with Bayesian Regularization was trained to predict motor angles from the 3D direction of a sound source. Experimental results demonstrate high accuracy, with standard deviations of 0.45° for flexion/extension and 1.02° for rotation, and mean estimation errors of 1.96° and 3.96°, respectively. The proposed CNN-based IK model enables smooth, precise head orientation despite biomechanical flexibility, offering a robust approach for humanoid robots requiring realistic motion in response to sensory cues.