Capítulo 4 Análisis de resultados
4.3 Brechas existentes
4.3.1 Análisis comparativo de brechas
Traditional scene analysis and object tracking relies mainly on data from vision cameras. To overcome limitations such as occlusion and a lack of 3D spatial information, numerous cameras, stereo-matching-algorithms and special depth measurement devices, such as the popular Microsoft Kinect [Kin], are commonly used. A tactile sensitive surface on the other hand provides an orthogonal approach to extend such scene analysis setups. A high- resolution tactile surface can supply information about the shape and position of multiple objects and also deliver information about the weight and pressure distribution of objects. To be able to augment different sized surfaces with tactile sensitivity, our main criteria for such a system was that it had to be modular, but at the same time simple to use and to reconfigure. After an exhaustive search for such a system, to the best of our knowledge no such system was available, and therefore we initiated our own development.
Figure 4.1: The resistance of a single resistive tactile sensor cell, measured between two electrodes, is the sum of sensor material volume resistance and contact resistances between the sensor material and the electrodes. The contact resistance changes according to the applied load on the sensor foam.
This project started out as a diploma project for Carsten Sch¨urmann [Sch08], who was
mainly supervised by me. The sensor developmental work was published at the World Haptics Conference in 2011 [SKHR11] and at Human-Robot Interaction conference in 2012, at a special workshop on Advances in Tactile Sensing and Touch [KS12].
4.1.1 Tactile sensing
The modular flat tactile sensor system implements a resistive tactile sensor array with a con- ductive foam used as the sensor material, similar to instrumented object from Section 3.3. Compared with numerous competing sensor technologies, such as broadly used capacitive or load cell based systems [ECL03,CMMS08,OBMC09], the resistive tactile cells are robust to electromagnetic interference, tolerant to overloading and require very few mechanical and electronic components making them easier to implement, especially for higher taxel counts. The applied pressure to output resistance has a hyperbolic characteristic, which enables high output resolution at low pressures, while simultaneously allowing a wide measurement range. Figure 4.1 depicts the components of the resistive tactile sensor cell’s resistance, Rt.
Rt is the sum of 3 parts – variable surface interface resistances Rs1+ Rs2 and a constant
sensor material volume resistance, Rv. Varying the applied load, varies the surface interface
resistance (Rs1& Rs2), thus allowing simple data acquisition with a constant pull up or pull
down resistor and an analog-digital-converter. In his diploma thesis, Carsten Sch¨urmann
optimized the shape of the tactile cell form (a sample of various shapes tested can be seen in Figure 4.2) and found the M-shaped cell to be most desirable as it was very responsive to input load changes and the output was least dependent on contact location inside one taxel.
A single tactile sensor module sized 80×80mm is composed of 3 layers [Figure 4.3a] - a databus layer (bottom), a sensor electrode layer (middle), and a conductive elastomer foam layer (top). The sensor electrode layer is made out of standard FR4 printed-circuit-board (PCB) material. The top layer consists of 16×16 M-shaped electrodes and on its underside 16 analog-digital-converters (of type AD7490) with 16-channels. The bottom databus layer implements a PIC32 microcontroller to collect the data from the top electrode layer. All
Figure 4.2: A sample of the tested resistive tactile sensor cell shapes. The middle “M” shape was experimentally found to perform best as it hat the most uniform sensitivity across a taxel. (Image source: [SKHR11])
four sides of the databus PCB have connectors that can be used for interconnecting further sensor modules to the system, allowing large 2D areas to be made tactile sensitive. The proprietary parallel bus has a bandwidth capability of 500,000 taxels/second, allowing up to 1.9kHz sampling to be achieved using a single sensor module (if the number of sensor modules on a bus is increased, the frame-rate drops accordingly). As master-controlled bus arbitration is used, there are no packet collisions, allowing all of available bandwidth to be utilized at all times. We have successfully evaluated such active table systems consisting of up to 6×6 sensor modules (480×480mm). Finally, the conductive elastomer foam on the top provides the necessary change in resistance according to applied load.
A single interface board based on an AVR32 microcontroller is connected to one of the corners of the databus modules and converts the data on the proprietary parallel bus to a standardized USB-Video-Class protocol. Typically used by USB-Webcams, this protocol allows tactile data to be delivered to a USB-host (e.g., PC) as a video stream, where each pixel corresponds to a pressure value of a taxel. An example of such a tactile image can be seen in Figure 4.3b, in which a right hand was placed on top of a tactile plate built from 3×3 sensor modules. It is also interesting to note that numerous algorithms from the computer vision domain can be used to process the tactile data. Furthermore, the USB- Webcam connection standard makes the tactile sensor usable with numerous closed source programs in the vision domain in which proprietary systems cannot as easily be connected.
4.1.2 Applications
As an initial sensitivity demonstrator of the system, software was developed to detect the location of the handle of a cup by just observing the tactile image [SKHR11]. The additional weight of the handle allowed immediate directional recognition due to the fact that more weight was distributed towards the handle. For all the cups tested, the directional prediction
error remained under 20◦.
In [SSS+09], Sch¨opfer et al. used two of the sensor modules to manipulate deformable
materials. Each of the modules were attached as a tool on the end of two 7 degrees-of- freedom robotic arms to form two movable tactile sensitive grippers. Using only tactile feedback, a Play-Doh compound was explored and successively deformed into a round ball.
In a similar robotic setup, Sch¨opfer et al. [SSPR10] took advantage of the high frame
rate of the modular flat sensors and used them to detect object slippage. Slip was detected using a Fast-Fourier-Transform (FFT) on the tactile data followed by a trained artificial neural network to detect specific characteristic slip patterns in the frequency domain, which even allowed a discrimination between different surface textures.
Sch¨urmann et al. [SKHR12] analyzed the end state anticipation on grasp control for two
(a) Single tactile sensor module with 16×16 taxels, measuring 80×80mm.
(b) Tactile image of a right hand on a 3×3 tactile module system.
Figure 4.3: Modular tactile sensor system with a high-speed USB-Video-Class interface. The sys- tem allows data to be captured at a rate of 500,000 taxels/second. For the case of a single module, this results in a frame rate of approximately 1.9kHz.
experiment, we built a small book sized tactile object using four of the sensor modules - two on each side, and added an additional 6D motion tracking sensor. The results clearly showed the different positioning strategies of the hand on the object in relation to the anticipated task. Depending on the task’s desired end state, participant’s chose postures from the beginning of the movement right through to the end of the movement that ensured the hand was in a comfortable position at the final goal state of the task (in psychology the term used for this is end state comfort [RvHC96]). Regarding the finger forces used by participants during the pick and place tasks, the experiment revealed no surprises - the results were in line with the expected forces required to maintain a steady grasp of the tactile book, showing slightly elevated forces during the acceleration phases.
An experiment by Naceri et al. [NME13], used a similar book-type arrangement of the tactile sensor modules, additionally attached on both long sides to two PHANToM [PHA] 6 degree-of-freedom haptic device simulators to investigate the strategy of humans in dis- tributing finger positions and forces under external force perturbations. The experiment revealed a substantial systematic variability of finger placement between participants, how- ever, within participants the placement was rather constant across numerous trials. More importantly the experiment revealed the fast haptic learning curve of the participants - all learned to compensate for the expected external perturbations after only the first trial.
Recently, Maycock, Essig, Schack and Ritter secured funding through CITEC, Bielefeld University, for a project, entitled Single and Dyadic Visuo-Haptic Task Learning to inves- tigate human haptics and other characteristics of participants as they solve maze game. The goal of the currently running project is to gain insights into how humans acquire a new manual skill. To achieve this, posture, eye-tracking and haptic data is to be gathered and analyzed. The tactile data is being gathered using four of the tactile sensor modules presented in this section. The modules are attached to the sides of the maze, two on each side measuring the top and bottom interaction forces [Figure 4.4]. The CAD development and construction of the tactile maze mechanics were done by me.
Figure 4.4: Maze with four modular tactile sensors, two on the left and two on the right (top and bottom), to capture the tactile interaction patterns while participants solve the maze. The inset shows the underlying CAD construction.