A SpiNNaker Application: Design, Implementation and Validation of SCPGs

A SpiNNaker Application: Design, Implementation and Validation of SCPGs Brayan Cuevas-Arteaga1, Juan Pedro Dominguez-Morales2, Horacio Rostro-Gonzalez1, Andres Espinal3, Angel F. Jimenez-Fernandez2, Francisco Gomez-Rodriguez2, and Alejandro Linares-Barranco2 1Department of Electronics, DICIS-University of Guanajuato, Carr. Salamanca-Valle de Santiago Km 3.5+1.8 Palo Blanco, 36885 Salamanca, Guanajuato, Mexico [email protected] 2Robotic and Technology of Computers Lab., University of Seville, Av. Reina Mercedes s/n, 41012 Sevilla, Spain [email protected] 3Department of Organizational Studies, DCEA-University of Guanajuato, Fraccionamiento 1, Col. El Establo S/N, 36250 Guanajuato, Guanajuato, Mexico Abstract. In this paper, we present the numerical results of the implementation of a Spiking Central Pattern Generator (SCPG) on a SpiNNaker board. The SCPG is a network of current-based leaky integrateand-fire (LIF) neurons, which generates periodic spike trains that correspond to different locomotion gaits (i.e. walk, trot, run). To generate such patterns, the SCPG has been configured with different topologies, and its parameters have been experimentally estimated. To validate our designs, we have implemented them on the SpiNNaker board using PyNN and we have embedded it on a hexapod robot. The system includes a Dynamic Vision Sensor system able to command a pattern to the robot depending on the frequency of the events fired. The more activity the DVS produces, the faster that the pattern that is commanded will be. Keywords: SCPGs ·Legged robots locomotion ·SpiNNaker ·Spiking neurons ·Hardware based implementations 1 Introduction Robotic locomotion is a highly active research field in artificial intelligence. Nowadays, several methods have been proposed to achieve locomotion in a variety of robots (i.e. wheeled robots, legged robots, swimming robots, flying robots, etc.); particularly for non-wheeled robots, bioinspired locomotion systems may be implemented [1]. These systems, commonly known as Central Pattern Generators (CPGs) imitate behaviors of biological neural mechanisms and they are capable of endogenously produce periodically rhythmic patterns to contribute in locomotion of living beings among other rhythmic activities such as digestion, swallowing, etc. [2]. However, they do not work isolatedly, since they interact with other parts of the central nervous system [3] and afferent sensory information may shape the CPG’s outputs [4]. CPG-based locomotion systems have several advantages over non-bioinspired ones, such as: rhytmicity,stability,adaptability and variety (see [5] for a detailed explanation of these features). These systems have been designed and implemented for biped [6,7], quadruped [8] and hexapod, [9,10] among other kinds of non-wheeled robots (see [1,2,5] for comprehensive reviews). The implementation of CPGs for either software or hardware applications involves previous phases of modeling, analysis and modulation, which comprises the kind of neuron model to use, the coupling and the structure of the connections in a network. The last deals with parameter tunning and gait transitions [5]. As mentioned before, an important aspect on the implementation of CPGs is the selection of a neuron model. In this regard, there are several neuron models with different degrees of plausibility. To date, spiking neurons are considered as the most plausible neuron models, they form Spiking Neural Networks (SNNs) which are considered the third generation of Artificial Neural Networks [11]; CPGs built as SNNs are known as Spiking Central Pattern Generators (SCPGs). In this paper a fully CPG-based locomotion system for the locomotion of a hexapod robot is proposed. Our proposal covers all phases suggested in [5]: modeling and analysis, modulation and implementation. The CPG-based locomotion system is conceptually based on works published in [8–10]; however we have changed the spiking neuron model and implementation platform by a more plausible one and a brain-like hardware platform respectively, i.e., the currentbased integrate-and-fire model and a SpiNNaker board. Three different gaits are generated by the locomotion system, and simulated and implemented on the SpiNNaker board and validated on a real hexapod robot. Moreover, our proposal incorporates gait transitions according to the activity sensed by a DVS camera. The rest of the paper is organized as follows: Sect. 2.1 provides information concerning the theoretical background around Spiking Central Pattern Generators. Section 2.2 provides information about the hardware used in this work, specifically on the SpiNNaker board. In Sect. 3, the design, implementation and validation of the system is explained in detail. Section 4presents numerical results of the simulation. In Sect. 5, we present a perspective of the work and we conclude in Sect. 6. 2 Materials and Methods 2.1 Spiking Central Pattern Generators SCPGs are a variation of the well-known and widely studied CPGs, which are specialized neural networks capable of endogenously produce rhythmic patterns. The CPGs contribute to living beings to perform actions without consent-effort such as walking among others [2,6]. Also, they have served as the basis of locomotion systems for non-wheeled robots with remarkable advantages over nonbioinspired locomotion systems [5]. SCPGs naturally handle spatiotemporal information as locomotion requires [10], which means that they receive and send information over time. In [8–10]are developed, implemented and tested SCPGs based on a biological study of insect locomotion [12] for legged robots by means of BMS neuron models [13]. The locomotion of legged robots is achieved in these works by means of spike-timeactivity where each servo-motor receives a spike train; the presence or absence of spikes indicates the state of a servo-motor at the current time. The SCPGs use the current-based leaky integrate-and-fire neuron model [14] as processing unit. Such model is one of the standard models in PyNN [15]. Equation (1) shows the equation of the model with fixed threshold and decayingexponential post-synaptic current, excitatory injection in Eq. (2) and inhibitory injection given by Eq. (3). dv dt =ie +ii +ioffset +iinj c m +vrest −v tau m(1) die dt =−ie tau syn E(2) dii dt =−ii tau syn I(3) where vstands for the current of membrane potential. The excitatory and inhibitory current injections are expressed by ie and ii respectively. ioffset represents a base input current added each timestep. iinj is an external current injection but in this case it is equal to zero. cmis the capacitance of the leaky integrate-and-fire neuron in nano-Farads. vreset is the voltage to set the neuron at immediately after a spike. tau mmeans the time-constant of the RC circuit, in milliseconds. tau syn Eand tau syn Iare the excitatory and inhibitory input current decay time constant respectively. The neuron model uses a tau refrac value for representing the refractory period, in milliseconds and finally, vthresh stands for the threshold voltage at which the neuron will spike. In PyNN the model is described as if curr exp and the code reads: eqs =brian.Equations( dv/dt =(ie +ii +ioffset+iinj)/cm+(vrest −v)/taum:mV die/dt =−ie/tausynE:nA dii/dt =−ii/tausynI:nA tausynE:ms tausynI:ms taum:ms cm:nF vrest :mV ioffset :nA iinj :nA  ) 2.2 Spiking Neural Network Architecture (SpiNNaker) To implement and validate our SCPGs we used a SpiNNaker board [16]. SpiNNaker is a massively-parallel multicore computing system designed for modeling very large spiking neural networks in real time. Both the system architecture and the design of the SpiNNaker chip have been developed by the Advanced Processor Technologies Research Group (APT) [17], which is based on the School of Computer Science at the University of Manchester. Each SpiNNaker chip consists of 18,200 MHz general-purpose ARM968 cores. The communication between them is done via packets carried by a custom interconnect fabric. The transmission of these packets is brokered entirely by hardware, giving the overall engine and extremely high bisection bandwidth. In this work, a SpiNNaker 102 machine is used. This board comprises 4 SpiNNaker chips and, hence, it has 72 ARM processor cores deployed as 4 monitor processors, 64 application cores and 4 spare cores. A 100 Mbps Ethernet connection is used as control and I/O interface between the computer and the SpiNNaker board. 5V-1A supply is required for this machine. This platform has been used in previous works by the authors [18,19], proving its robustness and versatility (Fig. 1). Fig. 1. SpiNNaker 102 machine. 2.3 Dynamic Vision Sensor The AER DVS128 retina chip (silicon retina) [20] consists of an array of autonomous pixels that respond to relative changes in light intensity in realtime by placing the address of that specific pixel in an arbitrary asynchronous bus. Only pixels that are stimulated by any change of lightning transmit their addresses (events are produced). Hence, scenarios with no motion do not generate output events. These addresses are called Address Event (AE) and contains the x and y coordinates of the pixel that produced the event. In this work, an AER DVS128 sensor was used to switch among the three different gaits, which has an array of 128 ×128 pixels. 7 bits are needed to encode each dimension of the array of pixels in this case. This Dynamic Vision Sensor also generates a polarity bit that represents the contrast change, where positive means a light increment and, negative, a light decrement. The DVS128 sensor is placed on the PAER interface that allows parallel AER through the CAVIAR connector [21]. 3 Design, Implementation and Validation 3.1 Design Three Spiking Central Pattern Generators (SCPGs) have been implemented into SpiNNaker. Each of them represents a different locomotion gait of the hexapod: walk, trot and run. These SCPGs were deployed in the machine using the PyNN library, which allows to easily create populations of neurons, connect them and assign weights to the connections. SCPGs need an initial potential stimulus to start running. In this case, the resting potential and the threshold of the neurons are set to −65 mV and −50 mV, respectively. Hence, an initial potential of −49 mV is set on the first population to make the SCPG start running at the beginning of the simulation. With this stimulus, the SCPG is able to run infinitely while providing the same output spike pattern. The current-based Leaky Integrate and Fire neuron model defined in Eq. (1) has been used in this work, and its configuration parameters are presented in Table 1. Table 1. Configuration parameters of the current-based LIF neuron model Parameter Value cm 0.25 tau m20.0 tau refrac 2.0 vreset −68.0 vrest −65.0 v thresh −50.0 tau syn E5.0 tau syn I5.0 ioffset 0.0 The SNN used to design each of the three locomotion sequences has basically the same architecture. The difference between them are the number of populations used, weights and delays for each of the sequences. It is necessary to mention that each of these sequences is implemented as a different SCPG. In Fig. 2, we show the spiking neural network topology. Here, red neurons represent the direct stimulus towards the servomotors of the hexapod robot, and the neurons in blue and yellow are used to balance the generation of the correct patterns in each case. As can be seen in the figure, blue connections represent excitatory activity and, on the other hand, yellow connections represent inhibitory activity. Fig. 2. SCPGs neural network architecture. (Color figure online) To create the complete design of a SCPG it was necessary to find the weight and delay values for each of the connections between populations of neurons. This was carried out using a exhaustive search method. At the end of the design of the SCPGs, a matrix is obtained, where each element of a column represents a servomotor and each row is an instant of time. Then, the absence or existence of a spike in this matrix will be translated to the movement of a specific servomotor on the hexapod robot. Figure 3shows the spike trains (SCPG’s output) representing the three different locomotion patterns that will perform the hexapod robot. 3.2 Implementation The implementation of the SCPG in SpiNNaker was done using the PyNN toolchain. The results are the three types of spike patterns required for the locomotion. These sequences are shown in Fig. 3. In each figure, X-axis represents the simulation time and Y-axis shows the twelve neurons required for the locomotion of a hexapod robot; one for each servo-motor. The neural activity of each neuron is transduced into electrical signals for moving its associated servo-motor. Specifically, locomotion corresponds to horizontal (coxa) and vertical (femur) movements, thus six neurons controls the horizontal movement and the other six the vertical one. The results were corroborated with the state of the art [9]. After verifying that the results were correct in simulation, we proceeded to implement a complete system to validate the implementation of the SCPGs, which is described in the following subsection. Fig. 3. Spike output patterns for each locomotion gait of the hexapod. 3.3 Validation In Fig. 4, a block diagram of our complete validation system can be observed. It consists of a DVS camera, which receives events according to a change of intensity either positive or negative. The information is sent as an AddressEvent-Representation (AER) format to an FPGA, which is used to synchronize and encode the incoming information in the specific format used by SpiNNaker (40 bits). The SpiNNaker is connected via Ethernet to a PC. For the mechanical validation of the SCPGs for this last, a hexapod robot was used, which consists of 12 servomotors: 6 to give movement to the extremities horizontally and 6 vertically. Each of the servomotors was connected to a PWM provided by an Arduino Mega board. Fig. 4. Block diagram of the complete validation system. The experiment consisted in connecting the complete system and, according to the frequency of events produced by the DVS, a specific locomotion sequence is executed. Different thresholds of event frequencies were set for this purpose. For example, if the DVS produces a few events, the robot would not perform any gait. However, if the DVS produces a moderate event frequency the hexapod would start walking, and so with trotting and running with higher event frequencies. All locomotion sequences were included in the same PyNN script. This scenario corroborates the operation of the complete system including both software and hardware. 4 Numerical Results In Fig. 5, we show the whole system running a locomotion pattern in real time. In such figure, we can easily identify the main elements of the system: a SpiNNaker board configured with a SCPG performing the trot pattern, which can be observed on the oscilloscope. Due to the fact that our oscilloscope can only register four analog signals simultaneously, we decided to register three signals at the same time, i.e. three for each of the movements: horizontal (coxa) and vertical (femur), making thus a total of 12 signals registered for each gait. To show the effectiveness on replicating locomotion patterns such as those found in vertebrates we performed numerical simulations for the three different (walk, trot and run) gaits using the SpiNNaker board and they are presented in Fig. 6. In such figure, the gaits are presented as follows: in the left side we can find the signals corresponding to the horizontal movements and in the right side for the vertical movements, both for the right legs performing the three gaits. The hexapod robot has a symmetric design. For this reason we only present the results for one of the hexapod sides (the legs on the right). Fig. 5. System configuration Fig. 6. Real time simulation 5 Stand-Alone Cognitive Locomotion System The work presented in this paper is part of a more ambitious project in which the non-wheeled locomotion robot is commanded by a SNN in real-time using neuromorphic sensors and cognition to react to different stimulus. For example, to make the robot to follow an object that produces higher event activity on the DVS sensor it can easily be discriminated not only the frequency of the events, but also by the position on the visual field (left, center or right) to command a pattern that makes the robot turn left or right [22]. The hexapod robot used in the previous section will host an FPGA-based platform that will command the actuators of the robot according to SpiNNaker output stream of events. The circuit on the FPGA will be collecting AER events from the DVS, converting them to 40-bit format in one part. Furthermore, it