Abstract—Structured Computational Polymers (SCP) is a concept of layered class of active material that can sense its environment and, due to its cognitive capabilities, react “intelligently” to those changes. In such a material, we envision semiconducting polymer based sensing, actuation, and information processing for on-board decision making to be combined into one active material. This paper describes incremental steps taken towards developing such a multifunctional active material, concentrating on distributed forms of actuation and cognition, with an intermediate goal of utilizing SCP as a “skin” of a soft robot – a robot, made of flexible materials, which is not bounded by its rigid structure and can adjust to its changing environment – with its sensing, cognition, and actuation embedded in the shape. We demonstrate, via experiment and rudimentary simulation, the feasibility of utilizing water hammer as a form of directed, distributed actuation. We also show that distributed form of cognition can be realized via a novel concept termed Synthetic Neural Network (SNN), which is a type of organic neuromorphic architecture modeled after Artificial Neural Network. SNN, based on a single-transistor-single-memristor-per-input for an individual neuron, can approximate the sigmoidal activation function with an accuracy of about 3%. A simulation of the SNN is shown to accurately predict the directionality of water hammer propulsion with an accuracy of 7.2 percent.
I. INTRODUCTION URRENT robotics systems have a limited capability to respond to contact (grasping, collision) with the
environment. Some of the current solutions include torque sensors in the joints, external imaging and range sensors [1]-[4]. Current state-of-the-art systems combine these sensors in order to appear “soft” upon interaction [5]. However, the performance is still far from ideal. Tactile sensors have been employed for a number of years, allowing the robot to crudely estimate the 3D geometry of an object that it interacts with [6]-[7]. However, oftentimes these systems do not possess the necessary data processing capabilities to adequately handle fragile and delicate objects. What is
Manuscript received February 8, 2010. This work was supported in part by the NSF Safety, Security, and Rescue Research Center and the NSF under grants IIS-0841483 and CNS-0923518.
R. A. Nawrocki is with Department of Computer Engineering, University of Denver, Denver, CO 80208 USA (phone: 303-871-6794; fax: 303-871-4405; e-mail: robert.nawrocki@ du.edu).
X. Yang is with Department of Computer Engineering, University of Denver, Denver, CO 80208 USA ([email protected]).
S. E. Shaheen is with Department of Physics and Astronomy, University of Denver, Denver, CO 80208 USA ([email protected]).
R. M. Voyles is with Department of Computer Engineering, University of Denver, Denver, CO 80208 USA ([email protected]).
needed is a truly distributed form of sensing along with local and “smart” actuation based on the locally-available information. This need was the main inspiration to our work.
Structured Computation Polymers, or SCM, is a concept of a layered class of active material that has the capacity to sense its immediate environment and process that data for intelligent actuation – actuation where the applied force is based on the type of material handled. Fabricated using a concept called shape-deposition manufacturing [8]-[9], which allows for iterative combination of dissimilar material addition and removal with the use of sacrificial support materials, SCP will comprise of self-contained cells each with their own sensing, cognition and actuation embedded in the shape.
SCP will operate in 1D (linear), 2D (planar) and eventually even 3D (physical volume) space. It will have the ability to sense forces acting upon it and appropriately reacting to them. SCP will be fit with on-board information processing allowing for intelligent actuation that can be customized based upon the application and the operational environment. Because grasping a tomato requires different operational mode than tightening a metal bolt, SCP will allow for creating smart robotic actuators with the capacity to adjust grasping force for different objects. Other possible applications, including coating bridges with SCP, will provide information about cracks or bulging of gusset plates allowing for proper authorities to be notified. Pilots flying planes covered with SCP will have information concerning bending of the plane’s fuselage at their fingertips. Other uses, such as smart walls, body armor, and smart clothes (gloves, shirts) are practically endless.
II. LONG-TERM GOALS
In order to provide a feasibility study of the aforementioned Structured Computational Polymers our group is involved in a project aiming to create a soft robot where the SCP will be used as a “skin” for active interaction with its environment. Soft robots are a new group of mechanisms where their structure is not rigid but can adjust to the changing environment [5]. Commonly these systems are made from flexible materials, such as polymers. Semiconducting, conducting, and ferroelectric polymers have been used as piezoelectric sensors for strain gauges as well as actuators for contraction and expansion [10]-[13]. Also, a number of research groups have demonstrated complete robotics systems based in polymers [14]-[15].
Structured Computational Polymers for a Soft Robot: Actuation and Cognition
Robert A. Nawrocki, Student Member, IEEE, Xiaoting Yang, Student Member, IEEE, Sean E. Shaheen, Richard M. Voyles, Senior Member, IEEE
C
2011 IEEE International Conference on Robotics and AutomationShanghai International Conference CenterMay 9-13, 2011, Shanghai, China
Yeom et al., in [14], demonstrated a biomimetic jellyfish robot created with an ionic polymer metal composite that mimics the real locomotive behavior of a jellyfish. Sameoto, in [15], created an all-polymer foot capable of climbing walls, mimicking the ability of a gecko or spider foot. Both of these designs were inspired by biological systems. Polymer-based chemicals used as fuel [16]-[17] were developed in the mid-1990s, though they haven't gained much commercial success.
We envision our soft robot to have its physical structure embedded in the shape of a tube, and propelled using a concept called water hammer (a liquid traveling through a pipe will exert a forward, jerky motion on the pipe, due to a sudden closure of a valve). As we will demonstrate, the direction of propulsion due to water hammer effect depends on the overall shape of the hose. Information from local bend sensors can then be used to change the overall direction of propulsion by affecting the shape of the tube via miniscule contractions or expansions of locally placed actuators.
To address these requirements SCP will comprise of multiple cells where each cell has the following capabilities: sensing (to extract the shape of the hose), cognition (to relate the shape to the desired direction of motion), and actuation (to obtained desired shape). Figure 1 demonstrates the concept of SCP. Sensing and actuation will be performed using polymer bend sensors and actuators. The data will be processed in a distributed fashion requiring distributed data processing algorithm as well as the capability for each cell to exchange the necessary information with other cells. For this data-processing task we believe that a form of a neural network is perfectly suited as it naturally allows for data processing to be distributed and, in the case of individual SCP cells becoming non-operational, it does not result in a catastrophic failure [18]. Coating polymer tubing with such active material as SCP will allow to prudently influence the shape of the tube based upon the current shape and the desired direction of propulsion, thereby utilizing the SCP’s one degree of sensing and actuating.
We note that recently a number of research groups have demonstrated successful creation of “electronic skin” [19]-[20]. However, the proposed designs concentrate only on the sensing ability of their material and do not address cognition nor actuation aspects. As such our proposal differs starkly with the aforementioned designs.
Figure 1: Proposed SCP design: Each cell, neighboring with 6 cells, will be equiped with a polymer micro sensor and actuator as well as distributed cognition (SNN), to relate the current and actuate the desired shape.
This paper addresses the all-polymer, distributed cognition aspect as well as the distributed actuation of the soft robot that the SCP will be embedded with. SCP polymer sensors and actuators, and wireless data communication are part of our long-term vision, but are not addressed in here.
III. COMPONENTS Our SCP will be used as a “skin” for our soft robot that
will allow it to intelligently interact to the changes to its environment. Each cell will be fit with small contracting/expanding actuators. However, the main source of directed propulsion for the soft robot will be achieved via a phenomenon known as water hammer.
A. Actuation – Water Hammer
The water hammer effect, also known as fluid hammer, has been known since the introduction of the modern plumbing. Until recently, thought, it was viewed as a negative effect with the capacity to destroy indoor and outdoor plumbing and, as a result, a number of remedies were developed to mitigate its effects. This phenomenon occurs when water traveling through a pipe experiences a rapid and sharp change in pressure usually facilitated by a fast closure of a valve. Increase in pressure, at the point of the closure, is brought about by the continuous motion of the flowing liquid. The intensity of the water hammer effect is inversely proportional to the time in which the valve is closed: the shorter the shutoff time the greater the force of the effect.
Recently Perrin et al. [21]-[22] demonstrated the feasibility of harnessing this potentially devastating effect towards a useful application. In their experiment a wheeled object was placed at the end of a tether which featured a looped hose with a shutoff valve, see Figure 2. The object was a remotely operated car with a small electric motor. Increasing the length of the tether resulted in increased weight which eventually presented a difficulty for the car to pull. The main challenge of the task occurred when the tether was placed in a situation that resulted in the tether being stuck. Following Perrin’s work, we conducted two experiments; one with the tether wrapped around two cylindrical objects, forming an S curve, and the second with the tether being stuck underneath a door. In both situations the jittery movement of the hose, caused by the water hammer effect, resulted in the tether being set loose enough for the small car to pull it.
A second set of experiments was aimed at identifying if the water hammer effect could be used as an exclusive source of actuation. The experiment demonstrated that, with a straight hose, the wheeled robot would be propelled directly along the length of the hose. We also conducted a drag test, in which a weight was placed on top of one tether with no water hammer and on top of a second tether aided by the water hammer effect. The conclusion was that the tether aided by the water hammer effect was able to pull a weight almost twice that of the unaided tether. For more in-depth experimental set up and results please see [23]-[24].
Figure 2: (A) Diagram of tethered water hammer design, with individual elements outlined. (B) Prototype of water hammer-assisted robot.
a) Computer Simulation We constructed a rudimentary computer simulation of the
resultant force vector due to the shape of the hose, acting on the front-mounted object. For the purposes of the simulation, the hose was considered to comprise of a finite number of elements, each in direct contact with adjacent elements. Each finite element had a point placed at its center. These individual points were connected, in the XY-Plane, resulting in an approximate representation of the shape based on the formed angles, measured in radians. These angles measured are analogous to the information collected by bend sensors placed along the length of the hose, which will be used to represent its shape. Figure 3 illustrates the concept.
Figure 3. Finite elements used to describe the serpentine shape of the hose, along with their corresponding Fx and Fy components. The measured angles are analogous to the information collected by the bend sensors.
For the 0th order of approximation, the angles (θ in Figure 3) were summed up resulting in unique angles for individual shapes. For the 1st order of approximation, the last element (the furthest away from the valve) was assigned X and Y components of the aforementioned angle. Each subsequent
element (closer to the valve) had its force calculated based on its X and Y components with a scaled force of the previous component added. The scaling factor d, left as a variable, was adjusted through various trials, between values of ‘0’ and ‘1’. Increasing the scaling factor corresponded to an increase in the influence of the previous component(s). Equations 1 and 2 show the formulae used to calculate the individual force components.
(1) sin (2)
Table I shows the obtained resultant vectors for 22 distinct shapes for both 0th and 1st order of approximation with varying d factor (values for resultant force vector are given in radians). Analysis of Table I reveals that increasing d from ‘0’ to ‘0.2’ does not produce a noticeable change in the vector. However, as d is increased beyond ‘0.5’, some of the shapes result in significantly varying vectors. This is greatly in line with the expectation that increasing the influence of individual elements will result in a greater change in the force acting on the front-mounted valve. Intuitively extreme values, d=0 and d=1, are non-realistic values and a value somewhere between those extremes should the most closely relate to real conditions.
TABLE I Resultant vectors for distinct shapes with different d values (computer simulation). Shape 0th
b) Experimental Setup The most realistic value of the d factor (from section
Computer Simulation), however, can only be evaluated empirically by comparing the results from the computer simulation to the results obtained in a practical experiment. Hence, the next thing that needed to be done was to experimentally obtain the data relating the shape of the hose to the force acting on the valve. We attached the valve to a stationary (mounted to a large metal plate) force sensor that would measure the force impacting on the valve through the water hammer, in the X and Y direction. The sensor
Shape of the hose
A
B
5117
registered the force in X, Y, and Z directions. We disregarded the Z direction as the tube existed in a planar space. For more details about the setup please see [23]-[24].
The experiment included measurement of 20 distinct shapes. For each of the shapes, the resultant force vector, obtained by extracting point(s) of the greatest magnitude (in XY plane) of all of the impacting forces recorded, was matched with 20 distinct points on the hose that were obtained from pictures taken of the shape (analogous to information collected by bend sensors) before the application of the water hammer (throughout the experiment the shape would slightly change due to the forces generated by the effect). Figure 4 demonstrates an example shape used for this experiment. Table II presents the data obtained, with 20 different shapes each with a distinct force vector (the values are given as the angle, in radians, calculated from the X axis in the counter-clockwise direction). Shape 0 is a reference shape, where the hose does not have any bends and the angle equals to π/2.
Figure 4. Setup of the water hammer experiment with force sensor mounted at the valve (XY coordinates in picture). Marks indicate points used to represent the shape of the hose; the first four yellow marks (light in black and white) are used to show the points used for shape extraction by neural network.
TABLE II Fitting resultant vectors (1st order approx. comp sim) with distinct d values with
force vectors obtained in water hammer experiment (WH Exp) with corresponding errors.
c) Comparison of Simulation and Experimental Data In order to verify the accuracy of our computer simulation
(see section Computer Simulation) we used the shape data obtained during the lab experiment and matched it with our resultant force vector from the simulation. This fitting process involved modifying the parameter d in Eq. 1 and 2 (the influence of individual finite elements on consecutive elements) until the difference was satisfactorily small.
Table II presents data calculated for three different values of d factor equal to ‘0’, ‘0.1’, and ‘0.5’. It can be seen that, for d=0.1, the average error is the smallest, or 7.2% (we also measured the error for values slightly higher and lower than ‘0.1’ but the error was greater in both cases). However, for d=0.5 error increases to about 10.7%. This finding indicates that the direction of propulsion is only mildly affected by the overall shape of the hose and the greatest influence is due to the direction or shape of the very end of the hose.
In the future we plan to investigate if our results would change when the water hammer effect is altered, either by increasing or decreasing it (facilitated by changes in the diameter of the hose or the water pressure).
To further verify the greater importance of points that are the closest to the front-mounted valve as opposed to points further away from the valve, we repeated the comparison of the computer simulation with the experimentally obtained values. This time we only included four points as opposed to 20 points. We compared information obtained from points numbered 1,2,3,4 with that of 1,3,5,7 and 1,5,9,14 (point no. 1 was the first point after the valve). The d factor was held fixed at ‘0.1’. Table III demonstrates our findings. It can be seen that spacing the points (equivalent with placing the bend sensors) closest to the front-mounted valve resulted in the closest approximation of the resultant force, or 7.4%. It should be noted that, even though the accuracy of simulation with four points is comparable to 20 points, accuracy suffers slightly when the overall number of measurements is reduced (7.4% vs 7.2%).
TABLE III Comparison of resultant vectors obtained with hose shapes resulting from denser points (1,2,3,4) and sparser points (1,5,9,14) – see Fig 4 for marking description.Shape WH
The conclusion of our experiments is that we have three different sources of information that all lead in the same direction, that the direction of propulsion due to water hammer effect is correlated with and can be extracted from the overall shape of the hose. This, we believe, validates our assumptions that the water hammer effect can be used, in concert with our proposed SCP, as a sole source of propulsion for a soft robot.
B. Cognition – Synthetic Neural Network
Having validated that the direction of propulsion can be extracted from the shape of the hose, we should be able to steer the robot (powered by the water hammer effect) by locally placed actuators. These, in turn, would either contact or expand based upon the received shape information and desired direction of motion. Therefore the SCP, or the “active/intelligent skin”, will be required to perform local, primitive cognitive tasks (relating current shape to the desired shape).
This approach, with the aforementioned cell-based architecture of SCP, necessarily requires a distributed computing architecture. For this task we have developed a prototype polymer-electronics based neuromorphic architecture, modeled after Artificial Neural Networks (ANN), termed Synthetic Neural Network (SNN). ANNs are usually employed in computation problems in which the input space and output data streams is extremely large and too complex to create a simple mapping. They have the additional advantage over serial computation of providing robustness against damage or removal of an individual cell, or a subset of such cells, which does not result in catastrophic failure but only a gradual reduction of the accuracy of prediction [18]. Because our SCP needs to possess the ability to map an unknown relationship, process the data in distributed fashion, and possess damage robustness, an architecture modeled after artificial neural networks seems to be a natural choice.
1) Prototype PCB Implementation for Water Hammer Direction Prediction
Before we implement our hardware neural network, we wanted to know if a conventional ANN would be able to learn the mapping of the shape of the hose to the resultant momentum vector. We trained a neural network in MATLAB (Neural Network Toolbox version 5.1), with a backpropagation algorithm, to a satisfactorily low error of about 2%. Because of our finding that only the points closest to the end of the valve have a noticeable effect on the direction of propulsion (section Comparison of Simulation and Experimental Data) a minimally-sized yet functional architecture for the network was determined to be 4 input neurons, 4 hidden neurons, and a single output neuron.
In order to test the concept of a fully functioning polymer-based SNN, we have first implemented the neural architecture using conventional PCB hardware. For this, an embedded system circuit with an ATMega 128 fit with a microcontroller was used, and software coding was done
with embedded C. The analog hyperbolic tangent sigmoidal activation function used by MATLAB was implemented as a piecewise linear approximation with 8 segments. To enhance the calculation efficiency of the board, we used the fixed point calculation. Our network was not capable of a live training but merely calculating the output of the network based on supplied input values and hard-coded connection weights manually exported from MATLAB. The data used for this training was a rough approximation of the shapes generated for water hammer experiment (section Experimental Setup). We verified that the output from our PCB-based NN matched that obtained in the software to the accuracy of about 6% (see Table IV).
TABLE IV Output of ANN train on MATLAB compared with output of ANN on PCB.
Having verified that both a conventional neural network software emulation and a PCB hardware emulation are capable of extrapolating the hose shape information and matching it with the directionality of the propulsion due to water hammer, we next discuss steps taken towards realization of the semiconducting polymer based SNN.
A number of research groups [25]-[27] have presented possible designs for realizing neural behavior utilizing transistors. However, most of these designs are either implemented using a large number of transistors, capacitors, current sources and other electrical elements or require an intensive and difficult assembly process [28]. Our approach is based on a much simpler design, motivated by a reduction of the complexity of both the individual neurons and the entire network.
We proposed [29] an architecture that is significantly less complex, with a smaller number of components, making it easier to create, occupying smaller physical space and less prone to manufacturing errors. Our architecture utilizes a single transistor, for summing functionality and sigmoidal output, and a single memristor per input to realize the multiplicative property of a synapse.
a) Single Neuron Functionality Equation 3 describes the operation commonly used to
compute the output of an artificial neuron. ∑ 3)
This function multiplies all the inputs (xi) by their corresponding connection weights (wi), sums all of these products (∑ ), and produces the output based upon the utilized activation function [29], denoted by K() in Equation
3. Therefore every representation of a neuron must perform multiplication, summation, and activation.
b) Memristor as a Synapse for Connection Weight
In a biological neuron the synapse is responsible for converting the chemical or electrical input signal and passing it onto the soma, or neuron body, for further processing [30]. Its purpose is to provide a neuron with an input that is proportional to the importance of the signal from that specific neuron: An input from an “important” neuron is passed much more easily to the neuron body than an input from an “un-important” neuron. It is generally accepted that the information is stored in the synaptic strength. Learning (neuro- or synaptic plasticity) is accomplished by modifying, either increasing or decreasing, the strength (sensitivity) of the synapse.
In an artificial neuron the functionality of a synapse is realized via a connection weight. Analogously to a biological synapse, connection weight from an “important” neuron is significantly greater than a connection weight from an “un-important” neuron. Learning in an ANN is accomplished by modifying the numerical value of the aforementioned connection weight.
Hewlett-Packard laboratory [32] was the first to announce the creation of a memristor based on an idea proposed by Chua in 1971. More recently work by the group of Likharev [33]-[34] has demonstrated the use of memristors to create variable strength connection weights in a network. A memristor (memory resistor) can be in one of two possible states; ON (low resistance, RON) or OFF (high resistance, ROFF). The operation of a memristor can be summarized as follows [35]-[36]. A memristor is in the OFF state (high resistance) until the input voltage is increased past a threshold voltage (VON). A memristor stays in the ON state (low resistance) until the voltage is reduced below a threshold voltage (VOFF). With the output of a memristor (current) equal to a product (see Single Neuron Functionality for the operation of a single neuron) of input (voltage) and connection weight (conductance), a memristor, with its bistable nature, seems a natural choice to realize the functionality of a binary synapse. Figure 5 portrays the IV characteristics of a memristor created in our laboratory, outlining the operation of the device as well as indicating the ON and OFF currents (that correspond to binary connection weights).
a) Transistor as a Soma for Summation and Activation Function
In both biological neurons and artificial neurons, the soma performs two functions: i) summation of the inputs (from dendrites in the biological case) and ii) mapping of the summed-input signal to the output signal, usually along a sigmoidal function. Also, an important property of a neuron is that there are minimum and maximum values that the output saturates to; for an unbounded input there is a bounded output.
Figure 5. IV characteristics of a memristor. A hysteresis curve can be seen: for V = 2 V two currents can be achieved for either ON (RON) or OFF (ROFF) states, ION and IOFF respectively. Numbers 1 – 5 indicate the direction of the scan.
We have developed a simple, single-transistor circuit that produces a suitable approximation to a neural soma. Given the weighted inputs described above, the single-transistor circuit sums the inputs and produces a nonlinearly scaled output that approximates a sigmoid. To emulate sigmoidal behavior, we used a single characteristic VSD curve by choosing VG based on utilized RON and ROFF memristive resistances. Additionally, our design allows different neural behaviors (for different types of neurons) to be achieved by employing different gate voltages. For an extended range of output values, VD could be made variable. Figure 6 shows the IV characteristics of an OFET developed in our laboratory.
Figure 6. OFET as a soma - IV characteristic curves of an OFET. Different Drain current curves can be seen showing the linear region of the transistor, which would be used as transfer function for the neuron. Only single IV characteristics would be utilized for artificial soma. However, different neurons (in different regions) can be made to operate with different Gate voltages.
3) Neuron Simulation All of the results and analyses presented here are based on
the electric characteristics of polymer-based devices fabricated in our laboratory. At present our memristor and OFET possess the necessary electrical characteristics to realize the individual SNN components. However, we note that substantial improvement in the processing and performance of the individual circuit components can be made. As work on these areas is underway, we present here the results based on these early, prototype devices that are nonetheless sufficient to demonstrate SNN behavior.
a) Single Synthetic Neuron Behavior Figure 7A shows a diagram that represents the design of a
single synthetic neuron with two inputs and a single output. As can be seen, for a binary connection weight, a single neuron consists of a single memristor per input, a single transistor and two resistors. A single memristor, in either the ON or OFF state, corresponding to either low (RON) or high (ROFF) resistance, is used as a binary connection weight. Figure 7B, depicting output voltage plotted against input voltage, reveals that the activation function of the synthetic neuron looks remarkably similar (differing by about 3%) to a sigmoidal activation function. It demonstrates that only a single transistor with one memristor per input is sufficient to obtain a sigmoidal activation function commonly associated with analog artificial neural networks.
Figure 7. (A) Schematics of a single neuron with two inputs (a binary connection weight, formed by a memristor, is represented by two resistors: RON and ROFF), denoted by Vin_1 and Vin_2, and one output, marked Vout, from an organic field-effect transistor (OFET) used for summation of memristive currents and to provide activation function. (B) Comparison of activation functions of a conventional neural network unit (MATLAB’s tansig was modified to decrease the upper and lower limits as well as its slope: y = 0.376 *(2 ./ (1 + exp(-1.5*x)) - 1)) with the synthetic neuron shown in Figure 7A with binary connection weights for output of a neuron (measured at a drain of an OFET) for input between -6V and 6 V. The difference between these two functions, (error=abs((abs(Matlab)-abs(SNN)/abs(Matlab))), was calculated to be approximately 3%.
a) Non-binary Connection Weights For some tasks binary connection weights are sufficient
for proper operation [33]. However, other tasks may require a finer granularity of the connection weights. We note that this can be easily accomplished by increasing the number of memristors assigned to an individual synapse. Three memristors, with only a single one being in an ON state, would result in a synapse with 4 possible values. However,
setting a subset of memristors to an ON state would result in 2n possible values, with n being the number of memristors employed per connection weight. Making memristors with variable ON and OFF resistances (accomplished by varying the size of the memristor) would result in more linear distribution of quantized connection weight values.
C. Water Hammer on SNN Simulation
Having validated the ability of software and hardware-emulated ANN to relate the shape of the hose to the directionality of the propulsion due to water hammer effect, we wanted to see if our SNN was able (preliminarily as a software simulation) to perform the same task. We manually exported the same weights as used for our ANN on PCB experiment (discussed in section Prototype PCB Implementation for Water Hammer Direction Prediction), to an SNN. Each synthetic neuron had inputs composed of three memristors which quantized the input weights to 8 levels (weight quantization error = 8.5%) used for quantized connection weights.
Figure 8 demonstrates 20 distinct propulsion directions (given in radians) associated with 20 distinct shapes, with outputs from ANN trained in MATLAB and SNN (the same shapes used for Tables II and III). Shape 0 (direction along a straight, or non-bent hose) is a reference vector (angle equal to 1.57 [rad]): shapes above or below the horizontal axis indicate shapes that resulted in angle of propulsion either to the left or to the right of the reference shape. It can be seen that our SNN is not only able to correctly classify two distinct patterns (left or right propulsion), but is remarkably accurate (average error = 6.7%) in relation to the ability to correctly identify the angle, making it suitable to perform the primitive cognitive tasks (mapping shape of the hose to the direction of propulsion) of the SCP. For more simulations on the performance of our SNN please see [29].
Figure 8. Comparison of outputs (given in radians) of ANN (blue or dark), train on MATLAB, and SNN (red or light) trained to detect a directed shape due to water hammer effect on a hose. Shape 0 was used as a reference point (see Neuron Simulation section for explanation). Horizontal axis (value of π/2) is used as the reference: values above the axis indicate vector to the left of the reference direction, and values below the axis indicate vector to the right of the reference direction.
IV. SUMMARY We have reported on several on-going efforts aimed at
development of an Structured Computational Polymers
(SCP) that will comprise of multiple cells where each cell has the sensing, cognition, and actuation capability. We reported on the progress on the feasibility study of the SCP in a form of an “intelligent skin” for an all-polymer soft robot with its physical structure embedded in the shape of a tube, and propelled using a concept called water hammer. We demonstrated that the direction of propulsion due to water hammer effect depends on the overall shape of the hose. SCP will be used to collect the information from local bend sensors and then change the overall direction of propulsion by modifying the shape of the tube via contractions or expansions of polymer actuators. The cognitive function of mapping the current shape to the desired shape of the tube will be performed, in a distributed fashion, via the concept termed Synthetic Neural Network (SNN), where each SCP cell will be fit with a single synthetic neuron. We presented the architecture and simulation of a single synthetic neuron, as well as simulation of operation of a network of synthetic neurons and their ability to perform mapping of the shape if the hose to the directed propulsion due to water hammer effect.
ACKNOWLEDGMENT This work was sponsored in part by NSF Safety, Security,
and Rescue Research Center and the NSF under grants IIS-0841483 and CNS-0923518. We would also like to acknowledge Rachelle Cobb from Rose-Hulman Institute of Technology for her work on the creation of an OFET.
References [1] T. A. Dwarakanath, B. Dasgupta, T. S. Mruthyunjaya, “Design and
development of a Stewart platform based force-torque sensor,” Mechatronics, Vol. 11, pp 793-809, 2001.
[2] K. Kaneko, K. Harada, F. Kanehiro, G. Miyamori, K. Akachi, “Humanoid Robot HRP-3”, Proc. of 2008 IEEE IROS.
[3] D. Fioravanti, B. Allotta, A. Rindi, “Image based visual servoing for robot positioning tasks,” Meccanica, Vol. 43, pp 291-305, 2008.
[4] J. B. Mbede, X. H. Huang, M. Wang, “Reflexive sonar-based navigation of intelligent robot manipulators in dynamic environments using fuzzy control,” Advanced Robotics, Vol. 14, pp 221-240, 2000.
[5] A. Albu-Schaffer, O. Eiberger, M. Grebenstein, S. Haddain, C. Ott, T. Wimbock, S. Wolf, G. Hirzinger, “Soft robotics,” IEEE Robotics & Automation Mag, Vol. 15, no. 3, pp. 20-30, 2008.
[6] B. J. Kane, M. R. Cutkosky, G. T. A. Kovacs, “A traction stress sensor array for use in high-resolution robotic tactile imaging,” J. of Microelectromech. Syst., Vol. 9, no. 4, pp. 425-434, 2000.
[7] E. J. Nicolson, R. S. Fearing, “The reliability of curvature estimates from linear elastic tactile sensors,” In Proc. of the IEEE Int. Conf. on Robotics and Automation, Nagoya, 1995.
[8] J. G. Cham, S. A. Bailey, J. E. Clark, R. J. Full, M. R. Cutkosky, "Fast and Robust: Hexapedal Robots via Shape Deposition Manufacturing," Int. Jou. Rob. Res., Vol. 21, pp 869-882, 2002.
[9] S. Kim, M. Spenko, S. Trujillo, B. Heyneman, V. Mattoli, M. R. Cutkosky, "Whole body adhesion: hierarchical, directional and distributed control of adhesive forces for a climbing robot," IEEE Int. Conf. Rob. Autom., 2007.
[10] J. N. Barisci, G. G. Wallace, M. K. Andrews, A. C. Partridge, P. D. Harris, “Conducting polymer sensors for monitoring aromatic hydrocarbons using an electronic nose,” Sensors and Actuators B: Chemical Vol 84, Issues 2-3, 15 May 2002, pp 252-257.
[11] Y. Osada, D. E. De Rossi, “Polymer Sensors and Actuators,” Springer, 2000, 448, pp. 110-114.
[12] K. Ikushima, S. John, K. Yokoyama, S. Nagamitsu, “A practical multilayered conducting polymer actuator with scalable work output,” Smart Materials & Structures, Vol. 18, Issue 9, Sept 2009.
[13] S. Maeda, Y. Hara, R. Yoshida, S. Hashimoto, "Self-Oscillating Gel Actuator for Chemical Robotics," Advanced Robotics, Vol. 22, pp 1329-1342, 2008.SW. Yeom, IK. Oh, “A biomimetic jellyfish robot based on ionic polymer metal composite actuators,” Smart Materials & Structures, Vol 18, Issue 8, Aug 2009.
[14] S.W. Yeom, I.K. Oh, “A biomimetic jellyfish robot based on ionic polymer metal composite actuators,” Smart Mater. Struct. Vol. 18, 2009.
[15] D. Sameoto, YS. Li, C. Menon, “Multi-scale Compliant Foot Designs and Fabrication for Use with a Spider-Inspired Climbing Robot,” Journal of Bionic Engineering, Vol. 5, Issue 3, Pg 189-196, Sept 2008.
[16] M. Nagao, S. Jayaram, M. Sugio, E. P. Waldi, “Studies on the dielectric strength of oil-polymer composite insulation under variable frequency AC voltages,” Electrical Insulation and Dielectric Phenomena, 1999 Ann Rep Con, Vol. 1, pp 309-312.
[17] M. B. Sheth, “Marine propulsion system with a polymer composite oil sump structure,” United States Patent 6913498.
[18] R. A. Nawrocki, M. Shaalan, R. M. Voyles, “Monitoring Artificial Neural Network Performance Degradation Under Network Damage,” to appear in proceedings of ANNIE 2010.
[19] S. C. B. Mannsfeld, B. C-K. Tee, R. M. Stoltenberg, C. V. H-H. Chen, S. Barman, B. V. O. Muir, A. N. Sokolov, C. Reese, Z. Bao, “Highly sensitive flexible pressure sensors with microstructured rubber dielectric layers,” Nature Materials, Pub. Online: 12 Sept 2010.
[20] K. Takei, T. Takahashi, J. C. Ho, H. Ko, A. G. Gillies, P. W. Leu, R. S. Fearing, A. Javey, “Nanowire active-matrix circuitry for low-voltage macroscale artificial skin,” Nature Materials, Pub. Online: 12 Sept 2010.
[21] D. P. Perrin, A. Know, R. D. Howe, “A Novel Actuated Tethered Design for Rescue Robots Using Hydraulic Transients,” 2004 IEEE International Conference on Robotics and Automation, 2004, pp 3482 - 3487 Vol.4.
[22] R. L. Feller, D. P. Perrin, R. D. Howe, “Validation and Explanation of Waterhammer-Based Locomotion,” Robotics and Automation, 2006, ICRA, pp 4264-4269.
[23] X. Yang, R. Voyles, K. Li, “Experimental Comparison of Robotics Locomotion with Passive Tether and Active Tether,” to appear in Proceedings of the 2009 IEEE International Workshop on Safety, Security, and Rescue Robotics, 2009.
[24] X. Yang, “Toward a Distributed Actuation and Cognition Means for a Miniature Soft Robot,” MS Thesis, University of Denver, 2010.
[25] R. Genov, G. Cauwenberghs, “Dynamic MOS Sigmoid Array Folding Analog-to-Digital Conversion,” IEEE Trans. On Circuits and Systems-I: Reg Papers, Vol 51, NO. 1, Jan 2004.
[26] H. B. Hammouda, M. Mhiri, Z. Gafsi, K. Besbes, “Neural-Based Models of Semiconductor Devices for SPICE Simulator,” American Jour. Of Appl Sci. 5 (4): 385-391, 2008.
[27] C. Rasche, R. J. Douglas, “Silicon Synaptic Conductances,” Jour. Of Comp. Neuro. Vol. 7, pp 33-39, 1999.
[28] F. Alibart, S. Pleutin, D. Guerin, “An Organic Nanoparticle Transistor Behaving as a Biological Spiking Synapse,” Adv. Funct. Mater., Vol. 20, pp 330-337, 2010.
[29] R. A. Nawrocki, R. M. Voyles, S. E. Shaheen, “Simulating Hardware Neural Networks with Organic Memristors and Organic Field Effect Transistors,” to appear in proceedings of ANNIE 2010.
[30] E. R. Kandel, J. H. Schwartz, T. M. Jessel, “Principles of Neural Science,” McGraw-Hill, pp 832-850, 2000.
[31] S. Haykin, “Neural Networks; A Comprehensive Foundation,” Prentice Hall, pp 392-404, 1999.
[32] D. B., Strukov, G. S. Snider, D. R. Stewart, R. S. Williams, “The Missing Memristor Found,” Nature, Vol. 453, pp 80-83, 2008.
[33] K. K. Likharev, A. Mayr, I. Muckra, Ö. Türel, “CrossNets; High-Performance Neuromorphic Architectures for CMOL Circuits,” Ann. N.Y. Acad. Sci. 1006, pp 146-163, 2003.
[34] Ö. Türel, J. H. Lee, X. Ma, K. K. Likharev, “Neuromorphic architectures for nanoelectronic circuits,” Int. J. Circ. Theor. Appl. 2004; 32:277-302.
[35] Z. Biolek, D. Biolek, V. Biolkova, “SPICE Model of Memristor with Nonlinear Dopant Drift,” Radioengineering, Vol. 18, NO. 2, June 2009.
[36] YB. Li, A. Sinitskii, J. M. Tour, “Electronic two-terminal bistable graphitic memories,” Nature Materials, Vol 7. Issue 12, Pg 966-971, Dec 2008.
