=Paper=
{{Paper
|id=Vol-2472/p8
|storemode=property
|title=A motor controller for TORVEbot based on a System On Chip
|pdfUrl=https://ceur-ws.org/Vol-2472/p8.pdf
|volume=Vol-2472
|authors=Aristide Emanuele Casucci,Mohamed El Arayshi,Khaled Issa,Lorenzo Canese,Paolo Lucantonio,Mario Patetta
}}
==A motor controller for TORVEbot based on a System On Chip==
https://ceur-ws.org/Vol-2472/p8.pdf
A motor controller for TORVEbot based on a
System On Chip
Aristide Emanuele Casucci, Mohamed El Arayshi, Khaled Issa, Lorenzo Canese, Paolo Lucantonio, Mario Patetta
Department of Electronic Engineering
University of Rome “Tor Vergata”, Rome, Italy 00133
Email: casucci.emyl@tiscali.it,elarayshi.mohamed@gmail.com,
khaledissa51@gmail.com, lorenzo@canese.it, paolo.lucantonio@outlook.it, mpatetta96@gmail.com
Abstract— This paper presents the implementation results of sensors without the necessity to redesign the electronic board
the motor controller for TORVEbot, a wheeled mobile robot during updates or for different tasks. The experience sharing
designed to be used in cooperative teams/swarms. The controller is obtained providing the robot of a Q-RTS module [4] and a
has been implemented on the XILINX System on Chip Zynq.The
Zynq SoC family integrates the software programmability of an wireless transceiver to communicate with the other robots in
ARM-based microprocessor with the hardware programmability the swarm.
of an FPGA, enabling key analytics and hardware acceleration Q-learning Real-Time Swarm (Q-RTS) is an iteration-based
while integrating CPU, DSP, ASSP, and mixed signal functionality reinforcement learning suitable for real-time systems. Rein-
on a single device. forcement learning (RL) is an artificial intelligence technique
The controller has been implemented on the Programmable
Logic and successively connected to the microprocessor using to train an agent to perform a task by interacting with the
the AXI interface. Results in terms of area occupation have environment. By trial-and-error, the agent adapts its behaviour
been presented. In order to verify the correct behavior of the through a rewarding mechanism, called reinforcement, which
controller, it has been integrated into TORVEbot. is a measure of its task-solving performance [5]
A crucial aspect of robots cooperating in a swarm is
the capability to adapt itself to dynamic environments. Such
capability is usually provided such robots also by Machine
I. I NTRODUCTION Learning (ML) algorithms. ML refers to the ability of comput-
Swarm Robotics consists of the use of several autonomous ers to learning from data. In the last few years, ML gained an
robots able to cooperate with each other to accomplish several important role in several fields that as health, computer vision,
tasks. Swarm robotics is the study of how to coordinate large and communications energy [7], [9], [11]. The availability of
groups of relatively simple robots through the use of local increasingly high computational power and the introduction of
rules. new technologies [12], [14], [13], [16], [17], [18], [19] have
Robot swarms find application in different fields as search increased the interest in ML.
and rescue, precision agriculture, military surveillance. In Thanks to ML robots are able to learn what to do when
this scenario cooperating in the swarm must be able to take analyzing data coming from sensors and consequently, they
decisions autonomously considering its state and the external are able to take decisions and be autonomous.
environment [1], [2]. In [3] the authors presented TORVEbot, In this paper, we present the hardware implementation of
a wheeled mobile robot designed to be used in cooperative TORVEbot motor control. The paper is organized as follow: In
teams/swarms. The main features of TORVEbot are: Sect. II the TORVEbot architecture is described, in Sect. III the
• Flexibility: The robot it is able to operate in different motor control system is discussed, in Sect. IV the experimental
environments and to accomplish different tasks results have been shown and finally, conclusions are discussed.
• Experience sharing: Robot units cooperating in the same
swarm are able to communicate with each other. The
capability to communicate with other robot units is very II. TORVE BOT ARCHITECTURE
important. It can be used to share experience and speed- TORVEbot is a wheeled mobile robot designed to work
up the accomplishing of the task [4] in cooperative teams/swarms developed at the University of
The flexibility is achieved providing the robot of reconfig- Rome Tor Vergata. The structure of the robot unit has been
urable electronic devices and of a common interface for all the built using a 3D printer with PLA material and is equipped
sensors the robot needs. The use of a common interface for all with the following devices:
the sensors, make possible to equip the robot of the appropriate • HC SR04 ultrasonic sensor with a range of 2 cm to 4 m
c 2019 for this paper by its authors. Use permitted under Creative and 1 cm resolution
Commons License Attribution 4.0 International (CC BY 4.0). • FS 90R servo motor
36
• A DILIGENT Pynq board that replaces the ARTY board
used in the previous version of TORVEbot
• DHT11 digital temperature and humidity sensor
• GY-BMP280-3.3 pressure sensor
• NRF24L01+ wireless transceiver
Fig. 2. The PYNQ board
1) FPGAs are able to execute matrix computation in a
very efficient way. ML algorithms as for example CNN,
SVM, SOM [8], [6], etc. are characterized by parallel
Fig. 1. The TORVEbot swarm robot unit: a) views of the CAD design; b) operations as vector matrix multiplications.
a built prototype
2) FPGAs consents efficient implementation of Ensemble
Machine Learning systems [10].
The use of reconfigurable electronic devices allows the
3) More ML algorithms can be executed in parallel being
possibility to reconfigure the robot for different tasks with fast
the FPGAs designed for parallel computing.
easy operation with no necessary changes of the hardware and
software as well as the corresponding mechanical design. The The Block Diagram of TORVEbot control unit is shown in
proposed TORVEbot robot is provided with XILINX PYNQ Fig.3.
board ( Fig. 2) equipped with a XILINX Zynq System On Chip Such as control unit is composed of the ARM CORTEX A
(SOC) composed on an ARM Cortex A Microprocessor and microprocessor and custom peripherals that are:
Field Programmable Gate Array (FPGA). The main features • A Machine Learning Accelerator.
of the board are: • Two PWM controllers for the wheels
• A 650MHz dual-core Cortex-A9 processor • Some Application Specific Sensors Interface ASSC. In
• DDR3 memory controller with 8 DMA channels and 4 our case, we have three interfaces for the HC SR04 , the
high performance AXI3 slave ports DHT11 and the GY-BMP280-3.3.
• High-bandwidth peripheral controllers: 1G Ethernet, USB • An interface for the wireless transceiver (Wireless Inter-
2.0, SDIO face).
• Low-bandwidth peripheral controller: SPI, UART, CAN, The microprocessor, communicate with the custom
I2C pheriperals using the AXI bus. AXI [30] is part of ARM
• Programmable from JTAG, Quad-SPI flash, and microSD AMBA, a family of micro controller buses introduced in 1996.
card The first version of AXI was first included in AMBA 3.0,
• Artix-7 family programmable logic released in 2003. AMBA 4.0, released in 2010, includes the
In addition to the reconfiguration capability, the FPGA second version of AXI, AXI4.
offers the possibility to efficiently execute algorithms that In the following, we focus on the motor control blocks (the
are characterized by a considerable level of parallelism and two PWM controllers for the wheels)
the possibility to execute several algorithms in parallel. This
solution allows the possibility to have a system composed of
a Microprocessor for the general purpose operations, as for
III. M OTOR C ONTROL
example, ADC and DAC interface and a Hardware Accelerator
for all those algorithms requiring parallel computing. The The robot motion represents a crucial aspect in robot design
use of mixed architecture composed of microprocessor and and for this reason, this issue is very discussed in the literature
a hardware accelerator is a common solution in the literature [28], [29]. TORVEbot is provided of the motor FS90R (Fig.
[21], [20], [22], [23], [24], [25]. These approaches are very 4). The FS90R is a micro servo designed specifically for
useful in the case of ML algorithms for three reasons, namely: continuous rotation. This servo can work with both 5 V and
37
Fig. 3. TORVEbot controller block diagram
controller has been connected to the ARM processor (by the
AXI LITE BUS) using the same toolchain.
IV. E XPERIMENTAL R ESULTS
After the VHDL description, synthesis and P&R have
been performed using the XILINX VIVADO 2018 toolchain.
Synthesis has been performed at 50 MHz (otiming constraint).
Implementation results are shown in Tab. I. Results refer
to a couple of PWM controller, one for the right wheel and
one for the left wheel. They have been used only 99 of the
53200 slice LUTs and only 231 of the 106400 slice registers.
The few hardware resources used for the implementation of the
Fig. 4. The FS90R
motor controller allows the possibility to have lots of hardware
resources available for the other block of the entire system
shown in Fig 3. The total on-chip power is 1.4 W
3.3 V servo signals. Such a motor can be controlled using
PWM signals having a control period of 20 ms.
The PWM controller block diagram is shown in Fig. 5 and TABLE I
it works in this way: The time is divided into temporal slots of R ESOURCE U TILIZATION
0.1 ms. At every 0.1 ms, the main counter increments its value
from 0 to 200 (20 ms is the total duration of the control cycle Resource Used Available
of the motor). The output PWN waveform is obtained by a SLICE LUTs 99 53200
comparator that compares the counter value with a threshold SLICE REGISTERs 231 106400
value provided at the input of the controller. If the counter
value is under the threshold the output is 1 otherwise it is The two implemented motor controllers have been con-
0. The threshold is provided to the controller by the ARM nected to the microprocessor among the AXI-lite interface.
processor through the AXI lite interface. Fig6 shows a screenshot of the entire project in the VIVADO
The controller has been simulated in the tool suite. The Figure shows the main block of the systems
MATLAB/SIMULINK environment (both floating point that are the ARM microprocessor, the system reset, the AXI
and fixed point simulation have been performed) and interface and the motor controller
successively it was coded in VHDL at RTL level and After the integration of the motor controller with the ARM
implemented using the XILINX Vivado toolchain. The processor, the controller has been tested in TORVEbot. Exper-
38
Fig. 5. Controller Simulink Block Diagram
iments have been performed programming the robot to execute R EFERENCES
some specific trajectories. The software for the controller has
[1] Seo, S., Yang, H., and Sim, K. ”Behavior learning and evolution of
been developed in C language in the XILINX VIVADO SDK swarm robot system for cooperative behavior” (2009). IEEE/ASME
environment. When the robot has to perform some specific International Conference on Advanced Intelligent Mechatronics, AIM,
movements, the microprocessor controls the custom peripheral 673-678. doi:10.1109/AIM.2009.5229933.
[2] Iaki Navarro and Fernando Mata ”An Introduction to Swarm Robotics”
implemented in the programmable logic using the AXI-LITE ISRN Robotics Volume 2013, Article ID 608164, 10 pages
interface. [3] Di Nunzio L. Cardarilli G.C., Ceccarelli M. Fazzolari R. ”Design and
requirements for a mobile robot for team cooperation” 6th IFToMM
International Symposium on Robotics and Mechatronics (ISRM 2019)
V. C ONCLUSION ARTICLE IN PRESS
[4] Matta, M., Cardarilli, G.C., Di Nunzio, L., Fazzolari, R., Giardino, D.,
In this paper we present the implementation results of the Re, M., Silvestri, F., Span, S. ”Q-RTS: A real-time swarm intelligence
motor controller for the TORVEbot, The controller has been based on multi-agent Q-learning” (2019) Electronics Letters, 55 (10), pp.
589-591.
implemented on the XILINX Zynq, composed of an ARM [5] Sutton, R.S., and Barto, A.G.: ”Reinforcement learning: an introduction”
microcontroller and a Field Programmable Gate Array. Im- (MIT Press, Cambridge, MA, USA, 2018)
plementation results show a very limited resources utilization [6] Cardarilli, G.C., Di Nunzio, L., Fazzolari, R., Re, M., Spano, S.
”AW-SOM, an Algorithm for High-speed Learning in Hardware Self-
(only 99 of the 53200 slice LUTs and only 231 of the 106400 Organizing Maps” (2019) IEEE Transactions on Circuits and Systems II:
slice registers) and a total on-chip power about 1.4 W. Express Briefs.
Other custom peripherals for the TORVEbot are under [7] Capizzi, G., and G. Tina. ”Long-term operation optimization of integrated
generation systems by fuzzy logic-based management.” Energy 32.7
development. In particular, we are working on the wireless (2007): 1047-1054
interface that will provide the robot of the NRF24L01+ digital [8] Giardino, D., Matta, M., Silvestri, F., Span, S., Trobiani, V. ”FPGA
transceiver. Such transceiver will be used by TORVEbot to implementation of hand-written number recognition based on CNN”
(2019) International Journal on Advanced Science, Engineering and
communicate with the other robots in the swarm and with a Information Technology, 9 (1), pp. 167-171.
data collection control unit. This unit will permit the remote [9] Capizzi, Giacomo, et al. ”Cascade feed forward neural network-based
monitoring of the robot. model for air pollutants evaluation of single monitoring stations in urban
areas.” International Journal of Electronics and Telecommunications 61.4
(2015): 327-332.
[10] Cardarilli, G.C., Di Nunzio, L., Fazzolari, R., Giardino, D., Matta,
M., Re, M., Silvestri, F.S. ”Span Efficient Ensemble Machine Learning
VI. ACKNOWLEDGMENTS implementation on FPGA using Partial Reconfiguration” (2019) Lecture
Notes in Electrical Engineering Springer, 2019
The authors would like to thank Xilinx Inc, for providing [11] Beritelli, Francesco, et al. ”Automatic heart activity diagnosis based
FPGA hardware and software tools by Xilinx University on Gram polynomials and probabilistic neural networks.” Biomedical
engineering letters 8.1 (2018): 77-85.
Program. [12] Susi, G., Toro, L.A., Canuet, L., Lpez, M.E., Maest, F., Mirasso, C.R.,
Pereda, E. ”A neuro-inspired system for online learning and recognition
of parallel spike trains, based on spike latency, and heterosynaptic STDP”
(2018) Frontiers in Neuroscience, 12, art. no. 780, .
39
Fig. 6. VIVADO block diagram
[13] Salerno, M., Susi, G., Cristini, A. ”Accurate latency characterization International Journal on Advanced Science, Engineering and Information
for very large asynchronous spiking neural networks” (2011) BIOIN- Technology 9.2 (2019): 677-684
FORMATICS 2011 - Proceedings of the International Conference on [26] Bracciale, L., Catini, A., Gentile, G., Loreti, P. ”Delay tolerant wireless
Bioinformatics Models, Methods and Algorithms, pp. 116-124. sensor network for animal monitoring: The Pink Iguana case” (2017)
[14] Yassin, I.M., Abdul Khalid, M.F., Herman, S.H., Pasya, I., Ab Wahab, Lecture Notes in Electrical Engineering, 429, pp. 18-26.
N., Awang, Z. Multi-Layer Perceptron (MLP)-based Nonlinear Auto- [27] Loreti, P., Catini, A., De Luca, M., Bracciale, L., Gentile, G. and Di
Regressive with Exogenous Inputs (NARX) stock forecasting model Natale, C., 2019. The Design of an Energy Harvesting Wireless Sensor
(2017) International Journal on Advanced Science, Engineering and Node for Tracking Pink Iguanas. Sensors, 19(5), p.985.
Information Technology, 7 (3), pp. 1098-1103. [28] Dewangga, S.A., Tjandrasa, H., Herumurti, D. ”Robot motion control
[15] Nasution, F.B.B., Bazin, N.E.N., Daliyusmanto, Zulfikar, A. Big data’s using the emotiv EPOC EEG system” (2018) Bulletin of Electrical
tools for internet data analytics: Modelling of system dynamics (2017) Engineering and Informatics, 7 (2), pp. 279-285.
International Journal on Advanced Science, Engineering and Information [29] Bracciale, L., Loreti, P., Detti, A., Paolillo, R., Melazzi, N. B. (2019).
Technology, 7 (3), pp. 745-753. Lightweight Named Object: an ICN-based Abstraction for IoT Device
[16] Susi, G., Cristini, A., Salerno, M. ”Path multimodality in a feedforward Programming and Management. IEEE Internet of Things Journal.
snn module, using lif with latency model” (2016) Neural Network World, [30] XILINX, IP documentation: AXI Reference guide, (2019).
26 (4), pp. 363-376.
[17] Coco, S., Laudani, A., Fulginei, F. R., Salvini, A., ”Bacterial chemotaxis
shape optimization of electromagnetic devices” (2014), Inverse Problems
in Science and Engineering, 22 (6), pp. 910-923, Taylor & Francis.
[18] Bonanno, F., Capizzi, G., Coco, S., Laudani, A., Lo Sciuto, G., ”A
coupled design optimization methodology for Li-ion batteries in electric
vehicle applications based on FEM and neural networks”, (2014) IEEE
International Symposium on Power Electronics, Electrical Drives, Au-
tomation and Motion, pp. 146-153.
[19] Bonanno, F., Capizzi, G., Coco, S., Napoli, C., Laudani, A., Sciuto,
G. L., ”Optimal thicknesses determination in a multilayer structure
to improve the SPP efficiency for photovoltaic devices by an hybrid
FEMcascade neural network based approach”, (2014) IEEE International
Symposium on Power Electronics, Electrical Drives, Automation and
Motion, pp. 355-362.
[20] Hauck, S., Fry, T.W., Hosler, M.M., Kao, J.P. The Chimaera Reconfig-
urable Functional Unit (2004) IEEE Transactions on Very Large Scale
Integration (VLSI) Systems, 12 (2), pp. 206-217.
[21] Cardarilli, G.C., Di Nunzio, L., Fazzolari, R., Re, M. ”Fine-grain re-
configurable functional unit for embedded processors” (2011) Conference
Record - Asilomar Conference on Signals, Systems and Computers, art.
no. 6190048, pp. 488-492.
[22] Schmit, H., Whelihan, D., Tsai, A., Moe, M., Levine, B., Taylor,
R.R. ”PipeRench: A virtualized programmable datapath in 0.18 Micron
technology” (2002) Proceedings of the Custom Integrated Circuits Con-
ference, pp. 63-66.
[23] Cardarilli, G.C., Di Nunzio, L., Fazzolari, R., Re, M. Cardarilli, G.C.,
Di Nunzio, L., Fazzolari, R., Re, M. ”TDES cryptography algorithm
acceleration using a reconfigurable functional unit” (2015) 2014 21st
IEEE International Conference on Electronics, Circuits and Systems,
ICECS 2014, art. no. 7050011, pp. 419-422.
[24] Razdan, Rahul, Brace, Karl, Smith, Michael D. ”PRISC software accel-
eration techniques” (1994) Proceedings - IEEE International Conference
on Computer Design: VLSI in Computers and Processors, pp. 145-149
[25] Ammendola, Roberto, and Pierpaolo Loreti. ”Design and Evaluation
of a Scalable Engine for 3D-FFT Computation in an FPGA Cluster.”
40