Tactile Sensor and Method for Operating a Tactile Sensor

Abstract

In an embodiment a tactile sensor includes a plurality of stress sensors and at least one contact body, wherein the stress sensors are configured to detect a load pattern applied on a detection surface of the contact body.

Claims

1.-10. (canceled)

11. A tactile sensor comprising: a plurality of stress sensors; and at least one contact body, wherein the stress sensors are configured to detect a load pattern applied on a detection surface of the contact body.

12. The tactile sensor according to claim 11, wherein the load pattern comprises a static tactile force and/or dynamic tactile force, and wherein the dynamic tactile force varies with a frequency between 1 Hz and 1000 Hz, inclusive.

13. The tactile sensor according to claim 11, wherein the stress sensors are integrated into a chip.

14. The tactile sensor according to claim 13, wherein the chip comprises a memory with a classification scheme, wherein the classification scheme assigns a set of output values of the plurality of stress sensors to a predefined load pattern.

15. The tactile sensor according to claim 14 wherein the classification scheme is generated by means of a machine learning algorithm.

16. A method for operating a tactile sensor comprising a contact body with a detection surface, a plurality of stress sensors and a chip, the method comprising: applying a load pattern to the detection surface of the contact body; transducing, by stress sensors, the load pattern to a set output values; and assigning, by the chip, the output values to a predefined class of load patterns by a classification scheme.

17. The method according to claim 16, wherein load patterns resulting from static and dynamic tactile forces on the detection surface are classified by the same classification scheme.

18. The method according to claim 16, wherein dynamic tactile forces are classified without spectral analysis of the output values.

19. A method for generating a classification scheme, the method comprising: predefining classes of load patterns; performing multiple representative measurements of each predefined class of load pattern with a reference tactile sensor; and define a decision tree ensemble by assigning sets of output values to each predefined class of load pattern respectively.

20. The method according to the claim 19, further comprising generating the classification scheme for a classification of load patterns resulting from static tactile forces and dynamic tactile forces.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Further advantages and advantageous embodiments and further developments of the tactile sensor and the method for operating the tactile sensor will become apparent from the following exemplary embodiments illustrated in conjunction with the figures.

[0031] FIG. 1 shows an exemplary embodiment of a tactile sensor and a chip;

[0032] FIG. 2 shows an exemplary embodiment of a method for fabrication of a tactile sensor;

[0033] FIGS. 3A and 3B show an exemplary embodiment and a schematic embodiment of a chip used in a tactile sensor;

[0034] FIGS. 4A and 4B show exemplary embodiments of stress sensors used in a tactile sensor;

[0035] FIGS. 5, 6 and 9 show exemplary embodiments of tactile sensors onto which a load pattern is applied;

[0036] FIGS. 7 and 8 show scatter plots of output values of exemplary stress sensors;

[0037] FIG. 10 shows scatter plots of the output values of exemplary stress sensors in response to load patterns with static and dynamic tactile forces;

[0038] FIG. 11 shows the output value of exemplary stress sensors and the classification which results from these output values;

[0039] FIGS. 12 and 13 show a 3D-plot of exemplary output value sets with respect to the chip surface;

[0040] FIG. 14 shows an exemplary measurement setup for generating the classification scheme;

[0041] FIG. 15 shows an exemplary output value of a single stress sensor for different load patterns; and

[0042] FIGS. 16A, 16B and 16C show exemplary implementation schemes of a decision tree ensemble utilized in a tactile sensor.

DETAILED DESCRIPTION OF ILLUSTRATIVE EMBODIMENTS

[0043] The same, similar or equivalent elements are provided in the figures with the same reference numerals. The figures and the proportions of the elements shown in the figures with each other are not to be considered to scale, unless units are expressly indicated. On the contrary, individual elements can be exaggerated in size for better presentation and/or better intelligibility.

[0044] FIG. 1 shows an exemplary embodiment of a tactile sensor 1 and an exemplary embodiment of a chip 10 utilized in a tactile sensor 1. The tactile sensor 1 comprises a contact body 200 into which the chip 10 is embedded. The mount 201 enables connecting the chip 10 electrically by means of a flex cable. In the present exemplary embodiment, the mount 201 is embedded in the contact body 200.

[0045] The chip 10 comprises multiple stress sensors 100.1, 100.2, 100.3, 100.4. The chip 10 comprises 32 stress sensors and all necessary readout circuitry. The chip is bonded onto a flexible flat cable for power supply and serial communication. Advantageously the chip is compact an easy to handle.

[0046] To form the tactile sensor 1, the chip 10 is embedded in the contact body 200 made of silicone (PDMS). The contact body 200 has the shape of a fingertip. Advantageously the chip does not require any additional electronic components and wiring.

[0047] Load patterns on the contact body 200, in particular the detection surface 200a, deform the contact body and induce stress profiles in the chip 10. The 32 stress sensors 100 are distributed over the chip area and allow measuring the stress distribution in the chip. The stress profiles are specific to direction and intensity of the load pattern.

[0048] A possible load pattern can be a dynamic tactile force, caused by an object 300 sliding along the detection surface 200a. The tactile sensor 1 is capable of measuring vibrations of up to 400 Hz, which is crucial for detecting such load pattern. Thus, the tactile sensor 1 has a sample rate of at least 960 Hz, in order to guarantees sufficient bandwidth. Since the stress distribution over the entire chip is measured, all 32 stress sensors are read, which reduces the sample rate to 30 Hz. The stress sensors are read serially, wherefore the vibration are still sampled with 960 Hz. Thus, the sample rate per stress sensor 100 corresponds to the sample rate per tactile sensor 1 divided by the number of stress sensors 100 per chip 10.

[0049] Vibrations are recorded at multiple points on the chip's 10 surface. The spatial distribution of the stress sensors 100 on the chip 10 has a neglectable effect on the measurement, as long as the wavelength of the vibrations measured is much larger than the distance of the stress sensors 100 on the chip 10. For example, the maximum distance of stress sensors 100 on the chip 10 is at least hundred times, preferably at least 1000 times, highly preferred at least 100000 times, smaller than the wavelength of a vibration to be measured.

[0050] FIG. 2 shows an exemplary method for fabrication of a tactile sensor 1. In a first method step a mold 203 is filled with a first silicone layer. A bottom surface of the silicone layer, facing the mold 203, becomes the detection surface 200a of the tactile sensor 1.

[0051] In a second step, the chip 10 is placed flat on top of the first silicone layer, on a side facing away from the mold 203. The silicone has sufficient viscosity to hold the chip in place.

[0052] In a third method step the mounting 201 is fixed to the mold 203.

[0053] In a fourth method step the contact body 200 is completed, by filling up the mold with silicone material. In this method step the chip 10 is completely embedded into the contact body 200. The mount 201 is partially embedded into the contact body 200.

[0054] In a fifth method step the contact body 200 is cured at 2 bar pressure and 60° C. The contact body 200 is made of PDMS (Dowsil 3140). The PDMS material advantageously provides tight mechanical coupling between the stress sensors 100 and the contact body 200.

[0055] In a sixth method step the mold 203 is removed from the contact body 200. After removal of the mold 203, the detection surface 200a is freely accessible.

[0056] FIGS. 3A and 3B show a top view and a schematic view of an exemplary chip 10 utilized in a tactile sensor 1. The chip is fabricated in a 0.35-μm process. The chip 10 comprises 24 transistor-based stress sensors 100 for measuring either in-plane shear stress of the difference of in-plane normal stresses are strategically distributed over the chip area. The stress sensor signals are processed by a variable-gain differential difference amplifier and digitized by a 10-bit SAR ADC. The chip 10 has a resolution down to 11 kPa at highest gain.

[0057] The 24 stress sensors 100 are consecutively connected to the readout unit by a 5-bit multiplexer (MUX). The processed stress sensor 100 can be biased in four directions by an additional multiplexer, which enables an offset-compensated operation due to a Wheatstone bridge architecture of the stress sensors. A variable-gain differential difference amplifier (DDA) together with a 10-bit successive approximation (SAR) analog-to-digital converter (ADC) performs the stress sensor readout.

[0058] As shown in FIG. 3A, the chip 10 has an area of 2×2.5 mm{circumflex over ( )}2.

[0059] The stress sensor 100 readout as shown in FIG. 3B is composed of a differential difference amplifier (DDA) with variable gain, followed by a SAR ADC and a gain controller (GC). Based on the digitized output value of the stress sensors, the gain controller determines the maximum ideal gain value with respect to the linear DDA output range on the basis of a binary tree search.

[0060] The architecture of the DDA offers one differential input pair for the stress sensors and a separate pair for the feedback. Thus, a high ohmic interface for the stress sensor connection is provided and impedance variations in the feedback path during gain variation are not influencing the load impedance of the stress sensor during readout.

[0061] FIGS. 4A and 4B show a top view of stress sensors 100. Each of the stress sensors forms a square of active elements and can be described by four transistors in a Wheatstone bridge arrangement. The orientation is indicated by the coordinate system (x,y) parallel to the chip 10 edges. The doping polarity has as strong impact on the piezo resistive coefficients of MOS inversion layers. For instance Πcan be ten times higher in p-type than in n-type silicon, while Π.sub.12 is nearly 50 times higher in the n-type silicon than in p-type silicon. The in plane shear stress is therefore measured by means of a NMOS type stress sensor 100 rotated by 45° with respect to (x,y). The PMOS type stress sensor is oriented parallel to the coordinate system. Thus, the PMOS type stress sensor is sensitive to the difference of in-plane normal stresses. All four transistors within each stress sensor have a W/L ratio of 5 μm/10 μm. These dimensions represent a good compromise between four main sensor characteristics.

[0062] Each NMOS stress sensor is activated by connecting the common gate of its transistor to the drain voltage. Each PMOS stress sensor is activated by connecting the common gate of its transistor to the source voltage. The power consumption of the overall system thus can be reduced, as only the processed sensor is activated temporarily.

[0063] FIGS. 5 and 6 show schematic side view and a top view of the tactile sensor 1 according to an exemplary embodiment. The tactile sensor 1 comprises a detection surface 200a, which is in direct contact to the object 300. A perpendicular static tactile force FS.sub.0, a forward static tactile force FS.sub.1, a backward static tactile force FS.sub.2 and sideward static tactile forces FS.sub.4, FS.sub.3 between the object 300 and the tactile sensor 1 are detectable load patterns.

[0064] FIGS. 7 and 8 show measurement results of load patters with static tactile forces in different directions as a scatter plot for four stress sensors. For this measurement class data is utilized to assign a measured set of output values to a predefined class of load patterns. Each point in the scatter plot represents on scatter sample and is color coded according to the load class it is assigned to.

[0065] Reference data, which is labeled “no contact”, is collected without any mechanical contact between the object 300 and the contact body 200. Data with pure normal static tactile force is labeled FS.sub.0. In addition to the normal static tactile force tangential static forces are applied by gradually moving the object 300 under the tactile sensor 1. The data is labeled FS.sub.1, FS.sub.2, FS.sub.3 and FS.sub.4 according to the corresponding direction shown in FIGS. 5 and 6.

[0066] The Classification scheme, in particular a decision tree ensemble, is trained to detect contact and classify the load pattern caused by the shear forces in different directions.

[0067] Depending on the stress sensor, some classes of load pattern result in similar output values of the stress sensor. However, the entirety output values, the set of output values, is distinguishable for different load patterns. The classification scheme considers the entire set of output values, whereby the applied load pattern is reliably assigned to a predefined class of load patterns. A trained random forest algorithm may assign the load patterns to the correct class of load patterns with 99.8% accuracy.

[0068] Furthermore, as shown in FIG. 9, a class of load patterns may be defined as the tactile sensor 1 sliding with its detection surface 200a along the object 300. The detection surface 200a may have a surface structure with multiple parallel ridges. The ridges cause a vibration of the contact body 200, when the object 300 slides along the detection surface 200a. Detection of the vibration enables to distinguish static tactile forces from dynamic tactile forces.

[0069] By measuring the vibration by means of a single stress sensor of the chip 10 at a sample rate of 960 Hz, it is possible to preclude, that the entire vibration is damped by the contact body. FIG. 15 shows the sensor output at three different states A, B and C. In state A the tactile sensor 1 does not move relatively to the object 300. However, in state A solely a static tactile force perpendicularly to the detection surface is present. In state B the tangential force is increased until the object 300 slides along the detection surface 200a. In state C the object 300 slips with 10 mm/s along the detection surface 200a and causes an oscillating output value of the stress sensor 100. Thus the dynamic tactile force can easily be detected by spectral analysis of the output value of a single sensor

[0070] By analyzing the output of multiple stress sensors 100, in particular all 32 stress sensors, both load pattern of static and dynamic tactile forces may be detected, without spectral analysis. FIG. 10 shows the complete classification scheme as a scatter plot for the two most significant stress sensors 100. Each point in the scatter plot represents one sensor sample and is color-coded according to the predefined load class assigned to the respective output value. The random forest classification algorithm is trained with at least 50 trees and unconstrained tree growth. The 5-fold cross-validation results in a total accuracy of 99.6%. Each class of load pattern reaches a true-positive rate of a least 98%.

[0071] There is no sensor dimension in which the set of output values in response to static tactile forces and dynamic tactile forces do not overlap. Thus, it is not trivial to reliably classify a load pattern by means of one or two stress sensors. A dynamic tactile force may only be classified by the relation output values of multiple stress sensors.

[0072] FIG. 12 shows a 3D-plot that visualizes output values of the stress sensors 100 over the chip 10 surface during static tactile force as load pattern. FIG. 13 shows a 3D-plot that shows output values of the stress sensors 100 over the chip 10 surface during dynamic tactile force as load pattern. For this visualization, the data is high-pass filtered to suppress any offset patterns from directional static force. Deviations from flat surface in the static plot are due to sensor noise. Hence, static tactile forces lead to well-defined stress patterns that can be distinguished from the distorted patterns in the slip sample. By training the classification algorithm may learn to distinguish these to patterns as two different classes of load patterns, whereby dynamic and static tactile forces are detectable by means of the same algorithm.

[0073] FIG. 11 shows a time domain measurement with the plurality of stress sensors 100 to validate the classification in a visual manner. The output values of four stress sensors are shown during the following load patterns: pure static normal tactile force FS.sub.0, static tangential tactile force FS.sub.1 or FS.sub.2, forward dynamic tactile force FD.sub.1. The order in the shown measurement is pure normal force for 2.5 seconds in the beginning of the measurement, static forward force is increased until 4.7 seconds of the measurement, the object slides along the detection surface 200a until 6.8 seconds, the sliding motion stops until 9.5 seconds, the sliding motion of the object along the detection surface 200a start again. The plot also shows the classification result for each sensor sample.

[0074] Whenever a sliding motion causes a vibration of the contact body 200. The vibration results in oscillating signals of the stress sensors. The classification scheme allows to detect and correctly classify the load pattern of the sliding motion mostly correct and reacts very fast to the sliding motion.

[0075] FIG. 14 shows a setup by means of which the classification scheme is generated. The setup comprises a classification mount 202 onto which the tactile sensor 1 is attached. By means of the classification mount 202 the tactile sensor is pushed against an object 300 with a definable normal tactile force FS.sub.0. The object 300 is a 3D-printed plate with different surface patterns. The surface patterns can be smooth, ridged, circles or triangles. The surface pattern allows to generalize the classification scheme from the surface. Otherwise the classification algorithm would be limited to classify load patterns of objects with only one specific surface structure. The object 300 is mounted on an X-Y-cross table. The object 300 can be moved in parallel to the detection surface 200a by means of stepper motors. The steps are fine enough to gradually increase the tangential tactile force without sliding the tactile sensor over the object. The stepper motors can accelerate and move the table with up to 10 mm/s to create slip. The tangential force which is applied to the detection surface is recorded by a verification sensor 5.

[0076] When generating the classification scheme, a predefined load pattern is applied to the tactile sensor 1 and the corresponding output values are assigned to the predefined class of load patterns. Such measurement is repeated for multiple times, to generate a reliable decision tree ensemble.

[0077] FIG. 16A shows an example of a single decision tree after training. In this context, the training is the measurement of predefined load patterns and the assignment of the corresponding output values to the class of load patterns. The features f.sub.i and threshold values t.sub.i of the nodes, the class labels C, of the leafs, and the entire tree structure changes during training. In the context of the tactile sensor 1, the stress sensors 100.1, 100.2, 100.3, 100.4 correspond to features f.sub.1, f.sub.2, f.sub.3, f.sub.4, the thresholds t.sub.i, correspond to output values of the stress sensors 100 and the classes C, correspond to predefined classes of load patterns.

[0078] FIG. 16B shows the static implementation in parts of a parallel architecture. In order to be able to continuously train the decision tree, the tree structure must remain reconfigurable.

[0079] As shown in FIG. 16C, the tree structure may be implemented by reconfigurable counter pails commonly used in FPGAs. The assignment of features to the nodes is achieved by a switching and interconnection network that wires the feature inputs to the appropriate node comparator. The Boolean function that encodes the tree structure and computes the leaf outputs is replaced by look-up-tables (LUTs) as used in FPGAs logic blocks. Only the comparators stay fixed and their thresholds must be stored in registers. This allows the architecture to support updates of the decision tree.

[0080] A full implementation of a decision tree ensemble comprises switching and interconnection blocks, which span the ensemble and allow comparators to be shared between trees. In particular, the trees of a decision tree ensemble may have different sizes.

[0081] The invention is not limited by the description based on the embodiments of these. Rather, the invention encompasses any novel feature as well as any combination of features, including in particular any combination of features in the claims, even if this feature or combination itself is not explicitly stated in the patent claims or exemplary embodiments.