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

1. Introduction

The Robotic Information System (RIS) is a novel information system design that uses intelligent robots to deliver data, information processing, and services to consumers across large geographic areas [1]. An abstract representation of the RIS shows three major pieces (Figure 1): database servers, the data and information processing engine (The-Oracle), and the head of a humanoid robot (Arslan). Each of those blocks was created with some enhancements to the current techniques that are typically employed to ensure prompt and effective response.

Figure 1. Robotic Information System (RIS) main blocks [1].

Figure 1. Robotic Information System (RIS) main blocks [1].

The primary front-end of the system is the Arslan humanoid robot, which has a 3D vision system [2], a 3D audition system [3], biomechanical systems for speaking and emotion, a smelling sensor, and an NLP system. Arslan itself contains a portion of the system’s AI, while The-Oracle contains the remainder.

In order to achieve a realistic motion, Arslan’s nick was developed to resemble a human’s nick structure using 3D-scanned vertebrae [4]. Since previous humanoid robots have been designed to execute rotation motions independent of displacement action, they lack this biomechanical design.

Since humans rarely do it in everyday life, Arslan did not need to complete every motion to the highest rating. Only flexion/extension and axial rotation have been permitted because lateral bending has very little bearing on Arslan’s application scope.

In a practical application, the Inverse Kinematics Model is necessary when Arslan detects a sound source and wishes to tilt its head in that direction. In order to achieve smooth motion of Arslan’s head, this study applies a motion dynamics model and uses CNN to predict those angles.

2. Methods

2.1. Arslan’s Nick Biomechanics

Axial rotation is permitted for C1-C2 because it has the biggest influence on axial rotation and to avoid conflict between flexion/extension motion and axial rotation at that level. Flexion/extension has been designed to be included in all levels from C2 to T1. Arslan’s cervical joints have been connected with springs and elastic bands, and two servo motors—one for flexion/extension motion and the other for axial rotation—have been used to control Arslan’s movement. Because of the realistic form of the nick, the head has been changed during flexion and extension, just like a human, and Arslan’s nick mechanism permits one side axial rotation of 41° and flexion/extension of 45°, giving it an appropriate range of motion (Figure 2).

Figure 2. Arslan’s Physical Structure and nick rotations [1].

Figure 2. Arslan’s Physical Structure and nick rotations [1].

2.2. Arslan’s Audition System

A computational model for sound source localization in 2D for the Arslan humanoid robot audition system using artificial neural networks has been introduced by Al Karaki [3]. Using two microphones, the system’s design aids in sound source angle prediction. The positive results confirm the effectiveness of the system’s approach and its feasibility for real-world use. Four microphones will be utilized in Arslan to pinpoint the sound source in three dimensions, and this two-dimensional model will be enhanced to incorporate the sound direction from two angles. Two microphones are situated in the ear canal, while the other two are situated on the forehead’s sides (Figure 3). A fifth microphone for Arslan’s speech audition is situated at its nose; it should be mentioned that those microphones are simply used to measure sound intensity.

Figure 3. Places of the microphones which are used in Sound source localization in Arslan humanoid robot.

Figure 3. Places of the microphones which are used in Sound source localization in Arslan humanoid robot.

2.3. Invers Kinematics Model Using Stereo Vision and CNN

Arslan’s nick is made up of seven joints that are held together by elastic bands and a spring, which prevents the joints from moving rigidly. Since only two motors control those joints, it is very challenging to determine the location and orientation of each joint because it depends on the motor angles, spring characteristics, spring tension, elastic band tension, the joints before and after it, and the momentum of various system components.

Arslan has a laser LED mounted on his nose that emits a beam that symbolizes the visual vector. As a reference, the laser spot location in 3D is transferred to a virtual location at the middle of the Arslan’s skull using a stereo vision system [5]. The real world angles (Theta and Phi) are computed by mapping vision vectors from the reference point to the laser spot at each setting over a range of angles applied to the flexion/extension motor (Th1 = 75 to 109 step 2) and head rotation (Th2 = 5 to 45 step 2). More than 2500 samples are produced by repeating this scanning process six times. Since the laser was invisible to the two cameras, some samples were omitted, making it impossible to determine the 3D location. To recover motor angles from speech source direction, a CNN with two hidden layers of 35 and 15 is employed, along with a training approach that uses Bayesian Regularization for 5000 epochs.

3. Results

The laser spot’s 3D placement at each motor’s angle pair is extremely near, despite the nick’s flexibility (Table 1). Given that the computed flexion/extension and head rotation angles had typical standard deviations of 0.45 and 1.02 degrees, respectively, sample scanning error can be disregarded (Table 2).

Table 1. Samples of data scanning results.

Motors Angles		Source Angles
Theta1	Theta2	Phi	Theta
75	5	76.13194	39.43156
75	7	75.93287	37.72342
75	9	77.54064	28.87504
75	11	78.12524	24.50305
75	13	78.55238	20.84372
75	15	78.92509	17.26678
75	17	79.2618	12.79393
75	19	79.53727	7.791852
75	21	79.76449	3.746861
75	23	80.09461	-2.45265
75	25	80.45018	-7.91286
75	27	80.5474	-13.9682
75	29	80.5205	-19.9437
75	31	80.36305	-25.3383
75	33	80.16096	-30.7363

Table 1. Samples of data scanning results.

Motors Angles		Source Angles
Theta1	Theta2	Phi	Theta
75	5	76.13194	39.43156
75	7	75.93287	37.72342
75	9	77.54064	28.87504
75	11	78.12524	24.50305
75	13	78.55238	20.84372
75	15	78.92509	17.26678
75	17	79.2618	12.79393
75	19	79.53727	7.791852
75	21	79.76449	3.746861
75	23	80.09461	-2.45265
75	25	80.45018	-7.91286
75	27	80.5474	-13.9682
75	29	80.5205	-19.9437
75	31	80.36305	-25.3383
75	33	80.16096	-30.7363

Table 2. Samples of data statistics.

Motors Angles		Source Angles
Theta1	Theta2	Phi	Th	AVG Phi	STD	AVG Th	STD
75	15	78.93	17.27	78.23	±0.39	16.05	±0.74
		78.36	16.44
		78.03	15.50
		78.05	15.72
		78.23	16.11
		77.80	15.24
75	17	79.26	12.79	78.59	±0.38	11.73	±0.59
		78.72	11.96
		78.40	11.37
		78.42	11.52
		78.59	11.64
		78.17	11.13
75	19	79.54	7.79	78.92	±0.35	6.85	±0.55
		79.06	7.02
		78.71	6.65
		78.76	6.67
		78.91	6.86
		78.52	6.12

Table 2. Samples of data statistics.

Motors Angles		Source Angles
Theta1	Theta2	Phi	Th	AVG Phi	STD	AVG Th	STD
75	15	78.93	17.27	78.23	±0.39	16.05	±0.74
		78.36	16.44
		78.03	15.50
		78.05	15.72
		78.23	16.11
		77.80	15.24
75	17	79.26	12.79	78.59	±0.38	11.73	±0.59
		78.72	11.96
		78.40	11.37
		78.42	11.52
		78.59	11.64
		78.17	11.13
75	19	79.54	7.79	78.92	±0.35	6.85	±0.55
		79.06	7.02
		78.71	6.65
		78.76	6.67
		78.91	6.86
		78.52	6.12

The CNN interprets motor angles as the network’s outputs and source angle values as the network’s inputs. To get better outcomes, inputs and outputs were standardized prior to training. The actual and anticipated outputs match well, according to the training results (Figure 4).

Figure 4. Training Results.

Figure 4. Training Results.

The average of the source angles is tested as the network’s input, and the output findings are compared to the angles of the associated motors for final verification. Theta1 and Theta2 have respective estimation errors of 1.96 (±4.46) and 3.96 (±1.88) for the average angles.

4. Conclusions

The inverse kinematics model developed for Arslan successfully integrates stereo vision data and CNN-based prediction to overcome the challenges posed by a biomechanically flexible neck structure. By replacing complex analytical calculations with a data-driven approach, the system reliably maps sound source directions to motor commands, achieving low estimation errors and smooth motion. The results validate the effectiveness of using deep learning for IK in robots with non-rigid biomechanics. This approach can be extended to other humanoid platforms where mechanical compliance and sensory integration are required. Future work will enhance the auditory system to full 3D localization, expand the dataset for more motion ranges, and investigate real-time performance in dynamic environments.