Abstract—High accuracy in modelling the behavior of human hand and fingers is obtained using control devices of high biological plausibility. Such devices are typically based on neural networks and are able to control in parallel multiple artificial muscles. This paper presents the structure of an electronic spiking neural network that was implemented to control the force of two opposing fingers of an anthropomorphic hand. In order to increase the level of bio-inspiration, the artificial muscles are implemented using shape memory alloy wires which actuates by contraction as the natural muscles. Moreover, the contraction force of the SMA actuators is directly related to the spiking frequency that is generated by the artificial neurons. The results show that using few excitatory and inhibitory neurons the neural network is able to set and regulate the contraction force of the SMA actuators. Index Terms—Force control, shape memory alloy, spiking neural networks, antropomorfic hand. I. INTRODUCTION Modelling the motor abilities of the human hand and fingers represents a challenging task for robotics due to smoothness and diversity of the natural motions. The design of the control devices for such robotic hands should model the behaviour of the motor neural areas (MNA) and their bidirectional communication with the muscles. The natural MNA stimulates the muscles through efferent neural pathways that includes the motor cortex and the central pattern generators. In the opposite direction, through afferent pathways, the MNA receives information from spindles about the muscle stretch during relaxation [1], and from Golgi tendon organs during contraction [2]. Considering that the frequency generated by the spindles increases with the muscle stretch by an external force [3], the spindles output can be used to determine the rotation angle of the articulation. However, this function cannot be applied when the muscle contracts because spindles response to acceleration dominate their response during the passive stretch [4]. When the muscles contract the Golgi organs respond to the force applied on the tendons providing information about the muscle activity [5]. Starting from this idea, in the sequel we will evaluate experimentally the ability of a biologically plausible structure of spiking neurons to control the contraction of artificial muscles. The neural network uses the output of a force sensor that provides information about muscle contraction as the Golgi organs. The spiking neurons represents the most accurate model of the natural neurons [6] and their implementation in 1 Manuscript received September 18, 2020; revised December 23, 2020. M. Hulea is with Faculty of Automatic Control and Computer Engineering, at Gheorghe Asachi Technical University of Iasi, Romania (e- mail: mhulea@tuiasi.ro). analogue hardware benefits from very fast response due to parallel operation of neurons, low power consumption and ability to process high complexity functions. The spiking neurons used in this work have these advantages and, being implemented on PCB hardware, makes the prototyping of the synaptic configuration easier. In order to achieve the smoothness and accuracy of the natural motions the artificial muscles should mimic the behaviour of the muscular fibres. Thus, in this work the artificial muscles are implemented with shape memory alloy (SMA) which actuates by contraction as the biological muscles [7]-[9]. Moreover, the contraction strength can be determined directly by the frequency of the electronic spiking neurons. The research done in this field, shows that the SMA actuators are suitable for actuation of bioinspired systems [9] starting from artificial fingers [10], insect legs [11] and wings [12] to an artificial jellyfish [13] and an anthropomorphic arm [14]. In SMA based applications the control of the contraction force of these actuators plays a critical role [15]–[17]. The precise control of SMA actuators force was performed using algorithms programmed on a microcontroller in a clamping vice [15]. Another method suitable for SMA control is represented by the neural networks (NNs) that were used to implement actuators for lightweight applications [16] including a SMA based endoscope [17]. The NNs are suitable, also, in robotics for the force control of different types of servomotor-based manipulators [18], [19]. In this work we used the newest class of NNs which represents the spiking neural networks (SNN) to control the contraction force of SMA wires that actuate two opposing fingers of an anthropomorphic hand. The earliest research that approached SNN to control the SMA actuators was performed by our group [20]. In this direction, we evaluated experimentally the ability of SNN to control the contraction of SMA in positioning of a robotic junction [21] and in laser spot tracking [22]. II. GENERAL CONCEPT The results reported previously [21] show that networks with few spiking neurons are able to control the rotation angle of a robotic joint when the mobile lever moves towards target positions. In that case the spiking neural network behaves as a regulator for the rotation angle even when the mobile lever is slightly loaded. A. Artificial Fingers Based on these observations we implemented and tested experimentally for this work a neural structure that behaves Force Control for Anthropomorphic Fingers Actuated by Shape Memory Alloy Wires Mircea Hulea 1 International Journal of Modeling and Optimization, Vol. 11, No. 2, May 2021 58 DOI: 10.7763/IJMO.2021.V11.778