Intelligent Senses in Robot Based on Neural Networks

Transcription

1 Intelligent Senses in Robot Base on Neural Networks Vimal Ganotra, Rashmi Sharma, Priti Dhingra, Sourabh Sharma, an Navin Kumar Abstract This Paper escribes a moel an an implementation of spiking neurons for embee microcontrollers with few bytes of memory an very low power consumption. The propose moel consists of an elementary neuron network that use Hebbian Learning to train a robot to respon to the environment implementing Artificial Intelligence (AI) in robot. The moel is implemente using ATMEGA8 Microcontroller base on AVR RISC Architecture an teste with an ability to move forwar an accoring to intensity of light without human intervention an external computers. Inex Terms AI, ATMEGA8 microcontroller, hebbian learning, robots. I. INTRODUCTION Now ays, automation is applie in every inustry. The human labor is being replace by robots, ex. CNC machines an PLC. The time is emaning more avancement in every sector of life. AI base on speech an image processing has been evelope at University of Emirates, UAE. The interest of man is increasing in AI an it has got no limits, weather it is replicating all the features of a human being or making a super human. In the recent past, with the improvement of the technologies associate with computing an robots, there has been a broa base attempt to buil emboie intelligences. But the peculiar nature of this fiel has resulte in the many attempts being almost entirely unconnecte. Because of the ifficulty an lack of success in builing physical robots, there has been a tenency towars computer simulation, terme Artificial General Intelligence where virtual agents in a virtual reality worl attempt to achieve intelligent behavior. By the 1980 s AI researchers were beginning to unerstan that creating artificial intelligence was a lot more complicate than first thought. Given this, Brooks came to believe that the way forwar in consciousness was for researchers to focus on creating iniviual moules base on ifferent aspects of the human brain, such as a planning moule, a memorymoule etc., which coul later be combine Manuscript receive May 16, 2010; revise June 12, V. Ganotra is with Department of Electronics in Vieocon, Karnal, Haryana, Inia ( vimalganotra007@gmail.com) R. Sharma is stuent in Department of Electronics an Communication Engineering, Maharshi Dayanan University, Inia ( exter2011@live.in) P. Dingra is with Research an Development team in Wipro Technologies, Inia ( priti.hingra4@gmail.com) S. Sharma is with Robotics Department in Digital Surveillance Inc, Noia, Inia ( saurabhfb@gmail.com) N. Kumar is stuent in Department of Electronics an Communication Engineering, Mahamaya University, Uttar Praesh, Inia ( kumarnavin.el@gmail.com) together to create intelligence [1]. Some evote Particular attention to the Question of what kin of basic element, possibly Corresponing to a moel of single neuron function, shoul be assume in neural Spatio neural network paraigm Network research. It is to this research on alternative basic elements for neural network moeling that we turn our attention first, because a principle is emerging from this research that may have important implications for our unerstaning of natural an perhaps artificial intelligence [2] Fig 1. shows the spatio-temporal neural network paraigm, a neural network interacts with its environment in a real-time, close loop fashion that is of funamental importance, with respect to the learning mechanisms that are employe, an perhaps with regar to the neural network's function, in general. Fig. 1. Spatio temporal neural network paraigm. The paper is organize into seven sections. Section 2 iscusses the term Hebbian learning. In section 3, Backgroun of this work, while section 4 presents the methoology an the propose architecture. Section 5 emonstrates the Evolution moel an implementation etails.section 6 presents the results. Finally, the paper is conclue in section 7. II. HEBBIAN LEARNING Our propose moel consiste of an elementary eight neuron network that use Hebbian Learning to train a robot to respon intelligently to input light stimuli. First, we ecie upon a task that woul accurately enote Hebbian learning. One of the most common examples of conitional learning such as Hebbian learning is seen in Pavlov s experiment with his og. In this experiment, when foo was offere to the og, it cause the og to salivate. At first the soun of a oorbell elicite no such response. However, Pavlov ecie to soun the bell when he offere foo to the og. After a few repetitions of this experiment, the og began to salivate at the soun of the bell even when no foo was present. Here foo was the unconitione stimulus, an the oorbell was the conitione stimulus. Similarly, in our work we first show that shining light in front of an or behin the robot elicits no response but pressing the push button causes the robot to move forwar or backwar. We 488

2 then press the button while shining the light on the robot an the neural network programme into the robot causes it to associate the light input with the push button input. Soon the robot moves forwar or backwar epening on whether the light is shine behin or in front of it in the absence of push button input. There are other neurons in this network that play an inhibitory role an prevent the robot from going too close to the light. They too isplay learning. Initially, the robot goes very close to a light source before it ecies to move in the opposite irection. As time passes by the robot gets more responsive to the light an oes not get too close to either light source. Fig 2. Shows Experiences change the way one perceives, performs, thinks an plan. They o so physically by changing the structure of the nervous system, alternating neural circuits that participate in perceiving, performing, thinking an planning. A very simplifie view of learning woul state that learning moulates (changes) the inputoutput, or stimulus action relationship of an organism. Certainly our environment influences how we react to it, an our reactions influence our environment. In orer to reach our en goal, we first programme a four neuron neural network in ATMEGA8 Micro Pushbuttons, with thorough testing at each level of complexity. Following this, we ae the harware interface. This involve integration of stepper motor control coe into our neural network such that the stepper motors woul step when the motor neurons fire. We then built the chassis an ae all the motors, LDR s, pushbuttons an Microcontroller boar to the esign. At the neural level, many ifferent types of changes can be imagine. For example, recorings from iniviual neurons in the hippocampus show that these neurons change their "place fiel" (i.e. responses to location in space) as the animal investigates the experiment. This change coul be ue to changes in the way visual stimuli affect these neurons (synaptic), or in the way the neurons respon to the same inputs (intrinsic). where Δw is the change in the synaptic weight connecting neuron j to neuron i an xi an xj are the activities (firing rates, action potentials) of neurons i an j, an µ is a scaling parameter often calle learning rate as shown in Fig 4. Fig. 4. Learning rate Hebb's moel, while it is a real-time learning mechanism moel, utilizes time in a way that may not be vali. Hebb assume that neurons correlate approximately simultaneous signals. Both the assumption of simultaneity an the assumption that signals are correlate by single neurons can be calle into question in that it has now been emonstrate that much stronger contact with the experimental results of classical conitioning stuies is obtaine if one aopts two alternative assumptions: first, that neurons are correlating sequential rather than simultaneous events as a basis for learning an secon, that neurons are correlating changes in signal levels (first erivatives with respect to time) rather than the signal levels, themselves. These Alternative assumptions yiel a real-time learning. III. BACKGROUND In previous work [3], the evolvability of spiking circuit architecture for vision base navigation of a mobile robot was investigate. The approach [4] escribe a large set of intelligent evices with embee microcontrollers(pic) with many ifferent memory for that reuce energy consumption (<1µA) between sensory upates instea of continuously upating neural Network. This paper [5] Introuce Kaburobo s rule, performance an implemente stock Robot such as machine learning in the artificial intelligence fiel. Fig. 2. Learning of organism. Fig. 3. Learning of neuron. A very simple example of learning at the organism level which has been worke out a neural level (Fig 3.) is that of sensitization of the gill withrawal reflex in Alyssa. In the sea mollusk aplysia, a light touch to the animal's siphon results in gill withrawal. Hebbian learning rule: The formulation of associative learning that has gathere the most attention for those stuying the brain was ue to Donal Hebb (see quote above). This proposition has le to a number of mathematical rules, the simplest of which is: Δwij = µ xi xj IV. METHODOLOGY In this work, the esign an Implementation of Intelligent Robot base on Neural Network using ATMEGA8 Microcontroller an a high performance approach is presente that has not been use in this manner before for such esign A. Applie Methoology The applie methoology is base on the ivie an conquers approach. Each block in the architecture was esigne an teste separately an later those blocks were assemble an extra moules were ae to compose the complete system. The esign targete the manatory blocks an features an the optional coing blocks were not inclue. B. Propose Work Fig 5. Shows the basic block iagram of propose Intelligent Robot base on Neural Network. The input is in 489

3 the forms of Light an push buttons, an controlling of movement of robot using programming in ATmega8 microcontroller. When we first shine light on the robot from ifferent irections it has no effect on the robot. We then press the button while shining the light from a particular irection on the robot an the neural network programme into the robot causes it to associate the light input with the push button input accoring to how it was traine initially. Say, every time we give light from forwar irection, we press left button. Soon the robot learns to move in the forwar irection when light is shine from the forwar irection. It is ecie to moel this type of system where there woul be multiple layers of Hebbian learning. Environment input in form of light flashe from a torch from any of 2 irections(forwar, backwar) Output was in form of motion in 2 irections. Fig. 5. Block Diagram of Intelligent robot Inhibitory action was har coe (to keep from strong light) A remote ha basic controls to move robot while training. Training: The input is provie in form of light an simultaneously it is move in a particular irection via remote to train the robot. Once traine, the robot can recognize light from ifferent irections an move accoring to how it was traine uring training perio apart from inhibiting strong light. Fig. 6 is showing the circuit iagram of intelligent robot implemente on ATMEGA8 microcontroller. are esigne, LDR usage, timers, LCD etc. After getting the motor control working, integrate this with the neural network programme to make sure that the network was working fine an showing the learning characteristics. Finally, time ecaying unlearning, inhibition features an a special function to show the arbitrariness of the simplifie output function an how powerful this learning technique is ae. Then, shine light from any of 4 irections an show that the robot moves an there exists both excitatory an inhibitory effects. There will also be real-time learning an unlearning as the robot is moving through its light course. Forwar Pushbutton: The input to this neuron is the push button press. It has an excitatory connection to the forwar motor neuron with a very large weight since it represents the unconitione stimulus to the network. Forwar Light: This takes light as input the voltage from the front light sensor. This has an excitatory connection to the forwar motor neuron an an inhibitory connection to the backwar motor neuron. This is because we ecie our neurons can t subtract input irectly, so instea we coe it so that the front sensor will excite the front motor neuron an inhibit the back motor neuron. The backwar pushbutton neuron, backwar light neuron, an the backwar shock light neurons all behave very similarly to the escriptions above but in the opposite irections. Finally, the forwar motor output an the backwar motor output neurons basically fire epening on which input neurons are firing an the strengths of those connections. After esigning the neural network, we ha to esign the robot an motor controls. The robot is controlle by two gear motors, an they step backwars or forwars epening on which neuron fires. We also ecie to buil a stan-alone MEGA8 prototype unit which woul just be on the robot. Furthermore, a very important part of the project was to emonstrate the learning process of the robot through the neural network. V. EVOLUTION MODEL AND IMPLEMENTATION Fig 7. Shows the propose moel of Intelligent Robot. The implementation use here, esigne to maximize exploration while preserving the best solution obtaine so far. Fig. 7. Moel of intelligent robot harware esign consist of Fig. 6. Circuit iagram of intelligent robot in learning moe. It was ecie to the esign process in many logical stages. The first stage: to get the harware working. Solering the PCBs an planne the I/O pins an other things. The simple programmers for remote control LCD LDR Remote Motor Driver Power Sources 490

4 DC +12 V Aapters use as power sources an 7805 was use to power the mcu. To rive the motors, L293 is use since the motors cannot be run irectly from microcontroller because it operates on 12 V an also mcu cannot hanle so much current. This evice, provies enough current an power to control the motor.l293 to control both the motors. The output from ATMEGA goes into the input of the L293. The outputs of the L293 are wire irectly into the motors. The following is a schematic of how to use the L293 to rive the motors. VI. RESULTS TABLE I shows the results of this propose approach with the results of Intelligent Robot with learning moe. If Forwar button is off, Reverse button is off for the same sensor values, there is no robot movement (enote in the table as XX) in any irection in the learning moe, when the learn button is off. If Forwar button is on, Reverse button is off for the same sensor values, the robot moves in forwar irection in the learning moe, when the learn button is off. If Forwar button is off, Reverse button is on for the same sensor values, the robot moves in irection in the learning moe, when the learn button is off. TABLE I: RESULTS OF INTELLIGENT ROBOT WITH LEARNING MODE SL.N O Learn butto n Forwar Learning moe Revers e sensor 1 sensor 2 Motion 1 OFF OFF OFF XX 2 OFF ON OFF Forwar 3 OFF OFF ON OFF OFF OFF XX 5 OFF OFF OFF XX 6 OFF ON OFF Forwar 7 OFF OFF ON OFF ON OFF Forwar 9 OFF OFF ON OFF OFF OFF XX Last row of table II is the learne moe, both the forwar an reverse buttons are off, only learn button is on, robot moves in forwar irection the sensor values are 90 an 980. Table III shows the Noe potentials in Learn an Learning an Learne moes. When the robot is in the learn moe the potential values at the four noes is zero. When the robot is in the Learning moe the noe values are 65, 73, 80, an 75. When the Robot is in the learne moe the noe potential value for noe 1 is 155 an other noes are having zero potential, this inicates the Robot is learne in the forwar irection. Fig 8. shows the noe potentials for learning moe for Forwar an irections. Fig 9. Shows the noe potentials for learne moe for Forwar irection. S.N O TABLE II: RESULTS OF INTELLIGENT ROBOT WITH LEARNED MODE Learn button Forwar Learne Moe Reverse Sensor 1 Sensor 2 Motion 1 ON ON OFF Forwar 2 ON ON OFF Forwar 3 ON ON OFF Forwar 4 ON OFF ON ON OFF ON ON OFF ON Forwar 7 ON OFF ON Forwar 8 ON OFF ON ON ON OFF Forwar 10 ON OFF OFF XX 11 ON OFF OFF Forwar TABLE III: RESULTS OF INTELLIGENT ROBOT WITH NODE POTENTIALS Noe Potentials S. NO Moe Noe1 Noe2 Noe3 Noe4 Motion 1 Learning Forwar 2 Learning Forwar 3 Learning Forwar 4 Learning Learning Learning Forwar 7 Learning Forwar 8 Learning Learning Forwar 10 Learning XX 11 Learne Forwar Potentials Fig. 8. Noe Potentials for Learning Moe Potentials Fig. 9. Noe potentials for learne moe NODE1 NODE2 NODE3 NODE4 NODE1 NODE2 NODE3 NODE4 491

5 VII. CONCLUSION The capability of esigning an implementing an Intelligent Robot base on Neural Networks is propose in this work. Base on above iscussion it can be conclue that an Intelligent Robot has been implemente an traine to transport things from one place to another. Goal is to esign a neural network that woul moel some real worl phenomena using microcontroller an to learn the usage of microcontrollers an some real wor electronics. It is ecie to use light an moel a moth s behavior via using a robot whose movement was controlle by the outputs of a neural network. Eventually the implementation of artificial intelligence in robot is stuie with the help of neural networks which can makes robots to be artificially intelligent. The esign implemente was targete to be mappe on ATMEGA8 Microcontroller. Thus we have shown that the propose esign of an Intelligent Robot was successfully implemente base on Neural Networks. In future work, the esign can be moifie an enhance further by aing two more LDR s an Push s an Robot can be mae to move left an right irection. Many extra features to the robot such as avoiing an obstacle carrying something. The robot woul seem to learn to avoi an obstacle an try imitating the sense organ of human beings an also try to implement various other learning abilities. REFERENCES [1] E. S. Brunette, R. C. Flemmer, an C. L. Flemmer, A Review of Artificial Intelligence, in Proc of the 4th International Conference on Autonomous Robots an Agents, Wellington, New Zealan, pp , [2] A. H. Klopf an J. S. Morgan, The Role of Time in Natural Intelligence: Implications for Neural Network an Artificial Intelligence Research, Air Force Wright Aeronautical Laboratories Wright-Patterson Air Force Base, Ohio [3] D. Floreano an C. Mattiussi, Evolution of Spiking Neural Controllers for Autonomous Vision-Base Robots, in Gomi, T, e, Evolutionary Robotics. From Intelligent Robotics to Arti cial Life. Tokyo: Springer Verlag. [4] D. Floreano, N. Schoeni, G. Caprari, an J. Blynel, Evolutionary Bits n Spikes, in Articial Life VIII, Stanish, Abbass, Beau (es) (MIT Press) pp , [5] J. Wang an F. L. Zhuang, Stuy an application of Stock Robot kaburobo base on Artificial Intelligence, presente at International Joint Conference on Artificial Intelligence, Vimal Ganotra born in Karnal, Haryana, Inia on 25 th July1989. He has obtaine his Degree in Electronics in 2011 from Kurukshetra University, Inia. Having interest in System Architectures, he is currently working with Vieocon in Inia. Rashmi Sharma born in New Delhi, Inia on 3 r December She has obtaine her Degree in Electronics & Communication in 2012 from Maharshi Dayanan University, Inia. Having interest in wireless communication an neural networks, she is currently working uner XBee base project. Priti Dhingra born in New Delhi, Inia on 4 th December She has obtaine her Degree in Electronics & Communication in 2011 from Kurukshetra University, Inia. Having interest in machine learning an bipe locomotion, she is currently working with Wipro Technologies, Inia. Sourabh Sharma born in New Delhi, Inia on 25 th July He has obtaine his Degree in Electronics & Communication in 2011 from Kurukshetra University, Inia. Having interest in Artificial Intelligence an machine learning. He has many national level awars in robotics. Currently working on government projects base on unerwater robotics. Navin Kumar born in New Delhi, Inia on 29 th October 1991.He is an unergrauate stuent of Mahamaya University, Inia. Having interest in robotics, he is currently working uner GPS navigation system project. 492