Tactile sensors can be based on different principles, including capacitive, resistive or optical methods, and are oriented to a broad range of applications, for instance in assistive or industrial robotics or rehabilitation and medicine in general [1]. The realization of the electronics depends on the specific approach and application, although there are a few major common concerns such as wiring, crosstalk or parasitic capacitors. Large tactile sensors with a high number of tactels (a tactel is a single sensing point on a tactile sensor array) and real-time operation are often required. Some pre-processing on the sensory plane results in a reduction of the amount of information to be transmitted to the central decision unit [2–10]. Moreover, detection and processing circuitry should be located near the sensor to avoid problems caused by long wiring runs. It also needs to have a low number of integrated circuits and I/O connections. This reduces the number of cables and allows it to be housed in hands and grippers.

Many tactile sensors have been implemented with technologies that allow the incorporation of circuitry on the same substrate [7–10]. These implementations achieve high spatial resolution, so they are suitable for applications such as Minimally Invasive Surgery (MIS), though they have also been developed for use in other environments such as industry. The circuitry implemented on the same substrate performs the signal conditioning. This is common in capacitive sensors because stray capacitors are a key issue, and amplifiers based on switched capacitors can be realized [8]. Certain preprocessing for sensors able to detect normal and shear forces is also realized on chip, as well as switches to address the array. More complex processing can also be carried out such as that for preprocessing of the tactile image based on convolutions [10]. Nevertheless, most tactile sensor systems or artificial skins are composed of patches that contain integrated circuits on a printed circuit board (PCB). Large areas can be covered with these patches that form a network of smart sensors and communicate with a central processing unit through a serial bus [3,4]. Different devices can be used as the core of the hardware in these patches as discussed in the following paragraphs.

ASICs to act as coprocessors for tactile sensors have been reported [5,6] and general purpose Integrated Circuits (ICs) have also been proposed for that task [11]. They can undertake error reduction, compensate for interference and convert analog-to-digital. It is the best choice in terms of area and power efficiency. Moreover, slippage detection in manipulative tasks with hands or grippers has to be done in the range of 2–4 ms. This means the whole array of force sensors of the tactile sensor (tactels) has to be processed in this time (analog-to-digital conversion plus detection algorithm). The high dynamic performance of ASICs allows the detection of slippage [6]. Unfortunately, ASICs are quite rigid and their programmability is low, so the possibility of updating their functionality once they are fabricated is limited.

Other implementations are based on microcontrollers [2–4]. This strategy usually requires a higher number of devices in the PCB board. This means that a large area is needed because of the space they take up as well as due to a more complex wiring. Furthermore, tactels are read and processed sequentially, hence the response time is poor and slippage detection for a piezoresistive sensor with a high number of tactels is not feasible. However, this approach allows the design to be updated, so the tasks to be carried out by the microcontroller can be changed simply by programming it again. Further improvements of the performance are achieved if the hardware is based on a PSoC. These devices have a set of analog and digital blocks to be configured by the user so hardware is reduced when compared to other standard microcontrollers. Nevertheless, the size of the array that can be addressed depends on the resources implemented on the PSoC and the number of input and output pins. Moreover, though the on-chip blocks allow some level of parallelism in the signal conditioning, programmed algorithms are executed in a sequential way.

The performance of a FPGA falls between these two previous strategies. They are flexible devices because they can be programmed, and at the same time they have a high dynamic performance due to the parallel processing they allow [12]. The main advantage of this strategy is the possibility of performing quite complex pre-processing in real time. As the system becomes more and more complex, many tactile sensors are used, for instance in fingers and palms, so the huge amount of data provided by these sensors should be pre-processed for the main controller to be able to manage it in real time. On the other hand, FPGAs do not commonly have analog-to-digital converters. Therefore, the use of external converters could increase the complexity and cost of the circuitry.

This paper demonstrates an implementation that does not need such external converters. It is based on the direct connection of sensors to microcontrollers [13]. Since the FPGAs have many I/O pins, they allow a very direct connection between the tactile sensor and the device. The smart sensor thus obtained is compact and powerful in terms of real time processing capability. This strategy was proposed by the authors in [14], where an implementation based on active integrators was also proposed to cancel crosstalk and cope with large array signal conditioning. An implementation of this circuitry for a specific raw sensor is presented in this paper. Moreover, signal conditioning circuits based on a PSoC and a microcontroller have also been made for the same sensor for the purpose of analysis and comparison. These circuits are based on ideas previously presented in conferences and journal briefs [15,16]. The comparison is to be taken as an approximation, since final performance depends on many factors, for instance the PCB technology, the encapsulation of integrated circuits and electronic components, or the specific device chosen for the electronics to be based on.

Many tactile sensors are made of sheets of piezoresistive materials. The sheet covers an array of electrodes and we obtain an array of force sensing resistors as tactels. However, if a continuous sheet is used, parasitic resistors are present between tactels in the array. Moreover, crosstalk is present even if the array is composed of discrete force sensors once they are arranged in rows and columns [16,17]. The reason is that the addressing tracks are shared by many tactels and they form resistive paths that cause crosstalk. A circuitry that does not cancel crosstalk is depicted in Figure 1(a), where an example of a parasitic path is also shown. Crosstalk due to the electrical addressing (there can be a certain crosstalk for mechanical reasons) is not registered if a single tactel is pressed. However, if a group of tactels are pressed crosstalk alters the readings for them. It can be hidden if the pressed area presents some symmetry with respect to rows or columns, but it is clear in the other cases. For instance, Figure 1(b) shows the output of a sensor with the signal conditioning in Figure 1(a) for a slash bar on it. Note that there is significant crosstalk. A few strategies to reduce the interference caused by these resistive paths have been proposed [18], although the best one that is commonly implemented is grounding. Its goal consists in having the same voltage at both sides of parasitic resistors, so they are virtually short-circuited. Circuits that implement this strategy will be shown in the next sections.

The local electronics for a tactile array of N columns and M rows based on a microcontroller PIC18F4680 (Microchip) are shown at Figure 2(a). These electronics are in charge of scanning the array, storing the data and sending it via CAN bus to a central processing unit. The latter has been a personal computer with a card to communicate with CANOpen for the results of this paper. This figure also illustrates the grounding strategy used to cancel crosstalk. Note that there is one operational amplifier per column in Figure 2(a). Their purpose is to set the voltage V_ref at the tracks of all columns. Since the voltage of all rows that do not contribute to the output is also set to V_ref, any possible parasitic path is short circuited. The output voltage of a column is given by: (1) V out = ( R G R ij + 1 ) V ref ,   V ref ≤ V out ≤ V DDwhere R_ij is the force dependant resistance of the element ij in the array, and R_G is the resistance to set the gain of the transresistance amplifiers at the output of every column. R G = ( V DD V ref − 1 ) R ij   min at Figure 2, where R_ijmin is the minimum value of R_ij determined by the pressure range of our application. Note that the output range is reduced by V_ref in this implementation. However, the A/D converter in the microcontroller can be configured to take V_ref as low voltage and still provide 10 bits if the excursion is higher than 3 volts. The operational amplifiers were chosen to be rail to rail to accomplish this, and the output can rise up to V_DD −40 mV@5 V@25 °C typically. The range in resistance is therefore: (2) R G V ref V DD − V ref ≤ R ij ≤ R G V ref V DD − V ref 2 NOBwhere NOB means number of bits.

The operational amplifiers should provide enough current; the maximum demanded current being V_ref/R_ijmin. The sourcing current of the LMV324 (Texas Instruments) is typically 80 mA for 5 V of supply voltage. Moreover, current switches in the analog multiplexor ADG734 (Analog Devices) must withstand a current N × V_ref/R_ijmin and perform as very low resistances when they are ON. Their typical resistance value is 2.5 Ω so the error introduced is very low. Another source of error is the noise in V_ref. The voltage reference REF3012 (Texas Instruments) with 0.2% accuracy is used to reduce this error. Note also that V_ref is a reference, not a voltage supply, since it does not provide any current in static condition, so the error it causes is negligible once the transitory is concluded. The input offset of the Op-Amp (operational amplifier) is very small, only 1.7 mV typically, so the error it causes is also small. Figure 2(b) shows the measurement of a few resistors obtained from the electronics in Figure 2(a) versus the value registered by a multimeter. Thirty samples were taken for every resistor value. If the electronics is designed to be connected to a CAN bus, a regulator is required to provide the low voltages for the microcontroller. The regulator should be able to provide a current of (V_ref/R_ijmin) × N plus the current demanded by the microcontroller and the other integrated circuits. A TL750L05 (Texas Instruments) was chosen for the results of this paper. All the integrated circuits were low power to reduce power consumption. Finally, the MCP2551 (Microchip) CAN bus transceiver was used to implement the interface to CAN bus. Please find performance data for this implementation in Section 7 and Table 1.

A more compact implementation can be achieved if the electronics is based on a Programmable System on Chip (PSoC) instead of a standard microcontroller. These devices implement more analog blocks that can be used in the acquisition circuitry. For instance, operational amplifiers in Figure 3(a) are implemented on-chip. The architecture in Figure 3(a) is similar to that in Figure 2 and Equation (1) is valid. Nevertheless, due to the limitations in on-chip resources some differences are observed. First, just four columns can be implemented. Second, rows are addressed directly by output pins of the PSoC whose drivers can only provide the voltage levels related to logical high and low values. The output that is set at logical low level drives the active row and its corresponding pin [R_wj in Figure 3(a)] sinks the current I_DRV. The remaining output pins are set at high impedance. In this way it is possible to implement grounding if a few bias resistors and an auxiliary voltage reference are used. They are R_bias and V_ref in Figure 3. Note that this bias forces a voltage V_ref at rows that are not driven and parasitic paths are short circuited because columns are also set to V_ref due to the negative feedback in the amplifiers. As in Figure 2, limited gain and offset of these amplifiers are sources of error. Typical input offset voltage is as low as 1.3 mV for the PSoC of this paper (Cypress CY8C29466), so the error it introduces is negligible. The amplifiers also have to provide the current sunk by the external circuitry. There is an analog buffer between the Op-Amp output and the resistance R_G in the feedback loop. This buffer is able to provide up to 40 mA.

V_ref at non inverting input of the Op-Amps is generated internally because of the lack of I/O pins to take this reference from an external source or to share it with the biasing resistor network. Therefore, an extra voltage source is required to provide the intermediate voltage (V_ref = 2.5 V). The latter can be obtained with a voltage regulator. Another alternative is the use of the reference V_ref generated on-chip, although one pin is dedicated to providing access to it and the number of columns is reduced by one (three with the PSoC of this paper). The mismatching between voltage levels of biasing sources is another source of error. The external reference is generated with the regulator LP2985 (Texas Instruments) for the results of this paper. Another important equation is: (3) V ref   ( 1 R bias + 4 R S   min ) ≤ I DRV maxto guarantee that the current sunk by the row driver does not exceed the limit imposed by the PSoC (I_DRVmax). For the device of this paper this current is 25 mA. The output impedance of the row driver introduces error too, similarly to the ON resistance of analog switches at Figure 2. The maximum voltage at this output for 25 mA is 0.75 V, so a linear approximation results in around 30 Ω of output impedance. Figure 3(b) shows the measurement of a few resistors obtained from the electronics in Figure 3(a) versus the value registered by a multimeter. Thirty samples were taken for every resistor value.

The operational amplifiers are rail to rail, which improves the dynamic range. However, it is important to say that this range is half that of the converters, because the range of the input is from V_DD/2 to V_DD. Therefore, Equation (2) is valid but NOB is one bit less than the resolution of the A/D converter on the PSoC. Finally, the PSoC implements resources to ease communication through serial bus (Universal Asynchronous Receiver-Transmitter, I2C and SPI). Please find performance data for this implementation in Section 7 and Table 1.

Figure 8 shows photographs of prototypes of electronics based on Microcontroller, PSoC and FPGA. They are connected to the same piezoresistive tactile sensor [20], although the PSoC is not able to scan the whole array (the detail in the figure corresponding to the PSoC shows an implementation where the PSoC is the only IC on the PCB and the electrodes are arranged to resemble the shape of a fingertip). The output of these sensors for different letter-shaped profiles on them is shown on the left of Figure 8. Moreover, Table 1 presents results obtained with these electronics, where the transceivers are not counted in the number of ICs required. As mentioned in the introduction, the comparison should be made with care because different alternatives could be followed in every case. Nevertheless, it provides interesting information for hardware designers.

Table 1 also includes data for another two proposals. First, the circuitry in Figure 2(a) was designed in a straightforward way and an analog multiplexor was chosen to drive the rows. However, it is possible to implement the strategy in Figure 3(a) and replace the multiplexor with an array of resistors. This would save four ICs and some area and power consumption. However, the inputs sink up to 8.5 mA for a voltage drop V_OL = 0.6 V. This means that there will be larger limitations in the range of resistance that can be measured and the errors due to the output impedance will be larger than those in the implementation of Figure 2(a). Since the PIC18F4680 has 11 A/D channels, it is also possible to take off the multiplexor MC14067 if the number of columns to address is lower than eleven. Note that this alternative reduces the number of integrated circuits to five without the transceiver. Second, the implementation in Section 5.2 was tested with a general purpose development card. Some data such as power or area consumption cannot be provided for this reason but other meaningful data is worth including in the table. The data marked with an asterisk are estimations from the datasheet.

The implementation based on the PIC has the best performance in resolution and power consumption. It is also able to scan quite large arrays of 16 × 16 tactels, although it is the less compact. The implementation based on the PSoC is very compact due to the resources implemented on chip and to the high driving capability of the I/O pins. The CY8C29446 has 12 configurable analog blocks. Many functions can be implemented with these blocks and are available as user modules like Delta-Sigma, Successive Approximation or Incremental Analog to Digital converters and programmable amplifiers. Moreover, the blocks can be configured by the designer if there is not a module in the library, which was the case in Figure 3(a). The dynamic performance can be improved if more than one converter is implemented to work in parallel. This is especially true for low resolution converters. For instance, the 14 × 4 array is scanned in 1.36 ms if four 6 bit successive approximation converters are used (the effective resolution is 5 bits, as explained in Section 4). The power consumption shown in Table 1 was measured when low value resistors of 1.68 KΩ where connected at the output to simulate the access to a row whose tactels have a high pressure on them. Note that the power consumption of the PSoC is similar to that shown by the PIC in absolute terms. However, the implementation based on PSoC consumes approximately five times per tactel more than the implementation based on PIC. This is due in part to the fact that V_ref is 2.5 V in Figure 3(a) while it is 1.3 V in Figure 2(a). The maximum resolution achieved is 8 bits to scan the 14 × 4 array in 31.8 ms, while the PIC scans 16 × 16 tactels in 15.6 ms with 10 bits of resolution.

Although a few tradeoffs between input range, resolution and speed are present, the implementation based on direct connection to an FPGA is the fastest. It was designed for a resolution of eight bits and an input range up to 8 KΩ. Larger input ranges can be achieved preserving the dynamic requirements with smaller capacitors, but effective resolution could be less for low resistance values corresponding to high pressures. The FPGA is more expensive than the PIC and the PSOC, and the PCB has four layers. The implementation with active integrators needs five ICs, a number higher that for those based on PSoC or on direct connection with passive integrators, but lower than for the implementation with the PIC. The power consumption per tactel is considerably lower than that with the PSoC and twice that obtained with the PIC. It also requires a four layer PCB. It is particularly interesting to note the possibility of implementing concurrency on the FPGA. This allows a very high dynamic performance to be obtained and algorithms with high requirements in terms of speed, such as slippage detection, can be dealt with [6,21].

Another test was carried out to examine the performance of the implemented circuits in the crosstalk cancelation. A resistor of high value was surrounded by resistors of low value in an array with shared tracks to address rows and columns that model a tactile sensor. This was considered preferable to making it with a tactile sensor to isolate crosstalk caused by electrical and mechanical sources. Crosstalk was not observed in any case, it was smaller than a Least Significant Bit. However, crosstalk was very significant if the electronics in Figure 1(a) was used instead of the others.

Finally, despite the circuitry being designed to minimize errors, it is worth highlighting here that many raw piezoresistive tactile sensors based on conductive rubbers or polymers show non linearity, hysteresis, drift and poor repeatability [20]. The powerful electronics based on FPGAs provides hardware able to carry out complex computations in a short time and effectively compensate these errors as much as possible. Moreover, despite the fact that the intelligence of the sensor can cancel many errors, the effective resolution of the resulting sensor will be less than that predicted without taking into account the limitations of the raw sensor. This could justify the choice of a circuitry with an analog to digital converter of lower resolution. However, another significant issue has to be observed. It is the nonlinear relationship between the force and the resistance the raw sensor shows. Moreover, note that the relationship between the output voltage (the input of the A/D converter) and the resistance R_ij is not linear in Equation (1). This nonlinearity can compensate that imposed by the raw sensor between force and resistance, but we will have different responses depending on the specific implementation [20]. Nevertheless, the non linear behaviour between the measured (the pressure) and the input of the A/D converter has to be taken into account to evaluate the impact of the limited resolution of the A/D conversion.

Some general conclusions may be drawn from the data obtained for the three discussed approaches, although they need to be qualified by other aspects as done in the previous section. The shortest scanning time is obtained by the implementation based on the FPGA, while that based on the PSoC obtains the longest and the PIC-based one gives a time that is in between and provides the highest resolution in number of bits. Nevertheless, very short scanning times can be achieved with the PSoC for low resolutions. Power consumption per tactel is also greater for the PSoC-based implementation than for the others, and the PIC-based implementation consumes the least. However, the most compact realization is achieved by the PSoC, although a smaller number of tactels can be addressed. The implementation based on FPGAs is also quite compact but it needs four-layer PCBs. This fact, together with the price of the main integrated circuit, makes the cost of the latter higher than the hardware based on the PIC or the PSoC. Finally, it is worth mentioning that the implementation based on FPGAs has a high degree of parallelism in the proposed analog-to-digital conversion and in the algorithms running on it, which is a key issue in real time operation.