Resistive flex sensors: a survey

With regards to the conductive component, the most commonly adopted one is based on carbon elements. We can refer to carbon powder [21–24], ink [25], particles [26], nanoparticles [21, 27], nanotubes [28], or carbon's allotropic form as a graphite [29, 30]. Transition metals can be used too, such as silver [31–35], copper [29], nickel [36], platinum [37], and palladium [38]. Others are electro-conductive rubber [39], carbon-loaded rubber [40], and some conductive polymers such as the ionic metal composite (IPMC) [41], and Poly(3,4-ethylenedioxythiophene) (or PEDOT or sometimes PEDT) [42], whose poor solubility can be partly circumvented with the PEDOT:PSS composite (figure 4) [43].

Zoom In Zoom Out Reset image size

Clearly, the adopted conductive material and the way it is layered determine the basic electrical properties of the sensor. In this way, sensor resistance is highly susceptible to any defects along the conductive surface. This is particularly significant when the conductive material is deposited not in the commonly adopted form of a thin film, but as sensor yarns (i.e. as in [44], where glass yarns coated with PEDOT:PSS were integrated within braided fabric).

The flexibility of the sensor mostly depends on the intrinsic properties of the substrate and its defect-free good adhesion with the conductive component. The substrate can be realized by means of different materials based on polymers, among which the most significant are polyvinylchloride (PVC) [32], polypropylene (PP) [25, 34], polypyrrole (PPy) [22], styrene-ethylene/propylene-styrene (SEPS) [34], and kynar (a brand name for polyvinylidene fluoride—PVDF) [25]. Among polymers, elastomers are adopted due to their elastic properties, so we have silicone rubber [45, 46], polyurethane (PU) rubber [47], and Evoprene™, a thermoplastic elastomer compound [21]. In addition, it is worth mentioning the polyester family, used by several research groups. Such a family includes polyethylene naphthalate (PEN) [29, 35], polymide (PI) [32, 35, 43], which is capable of thermal isolation [36], polydimethylsiloxane (PDMS) [48] of the silicones group, polyurethane [47] (a synthetic resin), Mylar (used to make heat-resistant plastic films), and the most common, polyethylene terephthalate (PET) [25, 32]. During recent years, a number of researchers have even adopted artificial tissues as substrates, so to make the flex sensors 'directly' wearable. We should also mention cotton [33], cotton/Lycra [49], Lycra [22], polyester/Lycra [26], Belltron®/Lycra® [50], tightly woven polyester, nylon fabric [21], and Spandex, which is a type of stretchy polyurethane fabric [31]. Unusual but interesting substrates are paper [30, 32], cardboard [30] or even human epidermis [51].

The electrical conductive element(s) can be engineered with the substrate according to different techniques, preferably low-cost to guarantee an affordable final device. Additive systems, headed by general printed electronic systems, are preferred, and strategically considered by several important companies such as BASF SE, Emagin Corp., Kovio Inc., Toppan Forms, Sipix Imaging Inc., Printechnologics GMBH, Polyic Gmbh & Co. Kg, Novaled AG, Nano Solar Inc., Aveso Displays, Blue Spark technologies, and E Ink Holdings Inc.

Fundamental printing methods are ink jet [35, 52], spray, screen [32], flexographic or gravure [53, 54] and transfer printing [55]. Other systems include photonic flash sintering [29], injection by air pulses [46], melt-mixing [21], percolation [56, 57], and electrochemical polymerization [43]. With respect to manufacturing standard methods, the aforementioned ones allow lower costs, shorter execution times, and the possibility of a pattern on both flat and even curved substrates, and with every sort of geometry, in an array [58] (figure 5(a)), matrix [59] or shaped form [60], or with any other distribution of conductive areas (figure 5(b)).

Zoom In Zoom Out Reset image size

RFSs can be made without any coating, known as base, bare or uncoated flex sensors, or a final protective coating can be realized by means of a dielectric epoxy coating [32], silicon rubber [27, 46] or simply adhesive rubber [25], polyester or polymide (Flexpoint Inc.). The coating material usefully provides an overall chemical and mechanical protection.

The concentration by weight of the conductive material with respect to the substrate, and its nature (powder, ink, paste or film), play key roles in the final electrical performances of the sensor. In the most common case of a conductive ink made of carbon particles in a binder, a cracked ink surface (figure 6) allows the cracks to open when placed in tension, due to the elongation of the material surface, causing the crack faces to separate and so increasing the resistance. When the material is returned to the flat position, the distance between the cracks' faces shrink and the resistance decreases. Cracks do not really repair themselves but their relative distance changes according to the bending angle; this makes the binder system in the ink critical, since if the bond is weak the ink flakes off the substrate and the sensor becomes useless.

Zoom In Zoom Out Reset image size

Without micro-cracks, the sensor offers better durability but lower resolution [61]. Given such behavior, the conductive material has to be placed upon the convex side of the bent surface, otherwise flexion would act in compression rather than in stretching. To obtain a bidirectional behavior, that is, the capability of the sensor to change its resistance when bent both back and forth, the conductive material has to be engineered on both surfaces of the substrate [62]. Or bidirectional sensing can be realized by two back-to-back flex sensors [63–66].

Given all the aforementioned characteristics, the electromechanical properties of RFSs can be tailored to meet the needs of many specific applications.

Originally, this type of sensor was developed by AGE Company (Abrams/Gentile Entertainment, New York) only and used in the so-called 'PowerGlove' (by Mattel), a sensory glove capable of measuring finger flexion. For some years there was no way of obtaining some of those sensors unless they were removed from the glove itself. Nowadays, RFSs have been commercialized by different brands, among which are Spectra Symbol (www.spectrasymbol.com, with 'FlexSensor4.5' commercialized by the online retail store Sparkfun.com), Abrams-Gentile (www.ageinc.com), Flexpoint Sensor Systems, Inc. (www.flexpoint.com, with the proprietary 'Bend Sensors®' carbon/polymer based ink) and Images SI Inc. (www.images.com, that furnishes the 'Bi-Flex Sensors™' capable of bidirectional behavior). The RFSs mostly commonly adopted and reported in the literature are those produced by Flexpoint Inc. and Image SI Inc. (figure 7).

Nowadays, the typical commercial RFS can be purchased for as little as tens of US dollars.

Zoom In Zoom Out Reset image size

An RFS is a passive device which does not require any bias or source power to work. When the sensor is flexed the substrate is consequently compressed and the conductive layer stretches, thereby increasing the overall resistance up to a maximum value corresponding to the maximum measurable angle of deflection. The electrical value can have a reactive (in particular capacitive) component, as previously demonstrated with an accurate RLC model (figure 8(a)), which can be developed to fit the electrical response of flex sensors under fast deformation movements. However, this reactive component has a negligible value (figure 8(b)) [67].

Zoom In Zoom Out Reset image size

Focusing on the real (resistive) component, an RFS can be characterized in terms of its resistance value R versus the flexing/bending angle ϑ (R versus ϑ), practically realizable by wrapping the RFS around a pivot with a fixed curvature radius [68–70] or on top of a varying-curvature surface [19, 71–73]. In general, the smaller the radius r of curvature and the more the involved conductive area of the sensor in the deflection is, the greater the RFS electrical resistance R will be. In order to give the full picture, it is also possible to relate the RFS resistance R versus the applied bending force F (R versus F) [74], or to the portion of the total length of the RFS (effective length, l) effectively subjected to the flexion/bending (R versus l) [75]. Attention has to be paid to drift problems, since an RFS can vary its resistance over time even when in mechanically stationary conditions [19, 71, 76].

Many authors utilize an RFS in series with a fixed resistor, thus forming a voltage divider, reporting RFS characteristics in terms of output voltage V (rather than resistance R) versus bending angle ϑ (V versus ϑ) [70, 76–81], and in this way strictly linking the result to the value of the fixed resistor, as investigated in [82].

The most interesting characterization relies on the 'R versus ϑ' curve which will be discussed further in section 3.1.2.

3.1.1. Characterization setups

3.1.2. Resistance profile and new setup

As already mentioned, measurements can be carried out by bending the RFS around a varying-radius pivot (e.g. cylinders with different radii), or around a fixed-radius pivot (e.g. a hinge).

In the first occurrence (pivot with a varying-radius r or, equivalently, curvature k = 1/r), different works demonstrated an exponential trend decreasing in resistance as the curvature k was increased (respectively k within: 0–25,4 mm⁻¹ [71]; 285–1500 mm⁻¹ [72]; 0–0.02 mm⁻¹ [73]), with negligible hysteresis. As a noticeable result, the work in [40] found that the 'R versus r' curve is not meaningfully affected if the RFS does not perfectly adhere to the pin around which it is wrapped during measurements, therefore the 'perfect adhesion' of the RFS on top of the joint being measured is not a mandatory requirement.

In the second occurrence (pivot with a fixed radius), the 'R versus ϑ' curve presents a nonlinear trend, but different authors report different types of nonlinearity. In [68], the nonlinearity was found to be polynomial, within the 0°–170° range, with very low sensitivity for angles of <45°. In [25], the authors found a second-order polynomial behavior within the 0°–100° range, without evidencing low sensitivity for any particular angular range. In [85], the nonlinearity curve fitted two different second-order polynomial curves for angles <30° and >30°, respectively, with very low sensitivity for angles <30°.

In order to take 'a step ahead' towards a reliable answer regarding the kind of nonlinearity, here we have meticulously carried out measurements on several RFSs, adopting a fully automated and highly suitable mechanical setup (figure 9). It was used to ten times, iteratively bending different RFSs inward (0°–120°) and backward (120°–0°)  to known angles and measuring resistance outcomes in 5° steps.

Zoom In Zoom Out Reset image size

The core of the setup was an anodized metal 0.1 μm-tolerance mechanical hinge, which allows the swinging motion of one leaf with respect to the other, fixed to a supporting base. The mobile leaf is welded to a cylindrical pin with the axis of rotation of the hinge. We utilized three tailor-made hinges differing in pin diameter (0.8 cm, 1.0 cm and 1.2 cm). The swinging motion was realized by means of a stepper motor (PD series by Trinamic, Hamburg, Germany) fixed to the cylindrical pin so that its rotation moves the unfixed leaf. The entire setup was fixed to a professional passive dumped optical bench (Newport Corporation, CA, United States) to avoid any vibration affecting the measurements [86].

Every 5° in bending, the swing was paused for 2 s, during which six resistance values were collected via a multimeter (34405A by Agilent, Santa Clara, CA, USA) and averaged to minimize noise errors. Both motor and multimeter were managed by a pc-running routine in Labview (system design software by National Instruments, Austin, TX, USA), in order to automatize the setup.

We carried out tests on several samples by Flexpoint Inc. and Images Co., with different lengths (1, 2, 3, and 4.5 inches), and which were differently overlaminated (polyester, polymide and no-overlaminated respectively).

The results obtained from sensors by Flexpoint demonstrated some variability even among sensors of the same type and length, but with a unique common trend. As an example, figure 10(a) shows the behavior of seven twin samples by Flexpoint. They were 2 inches long, not overlaminated, tested on a 0.8 cm pin, and range resistance differences among sensor results were evident. However, figure 10(b) shows a similar trend for the resistance values when normalized. Similar considerations result for measurements made around hinges with pins with different radii (0.8 cm, 1.0 cm and 1.2 cm, respectively), as the example of figure 10(c) shows.

Zoom In Zoom Out Reset image size

In all the occurrences, inward and backward resistance values succeeded in being superimposed, even after loading and unloading of the samples under test, in a way that that no hysteresis could be measured because they are 'embedded' within the 'natural' standard deviation of the measures. We found a slow but quite linear resistance increase for bending angles of less than about 30° (R² > 0.891), while a rapid and linear increase in resistance occurred for angles greater than 30° (R² > 0.994), with a sort of 'piecewise' behavior. However, the 'R versus ϑ' curve of all the RFSs by Flexpoint can be usefully approximated by a quadratic polynomial curve (R² > 0.999). In a recent work we demonstrated how a nonlinear 'R versus ϑ' curve behavior of RFSs cannot depend on the mechanical features of the RFS but strictly depend on the electrical ones [87].

Interestingly, there are some solutions which can be implemented to linearize the nonlinear behavior, at least within the 0°–90° range, as realized in three different works. The first solution consists of fixing a thin polyvinyl chloride (PVC) foil over the sensor carbon layer and realizing a current divider with a parallel resistance [88], leading to a linear regression coefficient of R² equal to 0.9915. The second solution consists of cutting the original rectangular sensitive part of the sensor to obtain a triangular shape, which allowed a linear regression value of R² up to 0.9893 [86]. The third solution considers other shapes of the RFSs, rather than the standard rectangular one, determining the kind of strip geometry which can lead to the maximum linearization [89].

In contrast to the RFSs by Flexpoint, the RFSs by Images demonstrated an inborn linearity, with R² up to 0.997 (figure 8(d)). However, this advantage is paid for in terms of a lower sensitivity since the resistance roughly 'only' doubles going from flat to a 90° bend angle, compared to the resistance of the RFSs by Flexpoint, which varies up to thirty times. As a result, the RFSs by Flexpoint are better suited to applications which require higher sensitivity, and RFSs by Image are better suited to applications where the linearity is mandatory.

Only one work investigated sensor resistance variation versus applied force, obtaining a 'R versus F' curve [74]. It realized flow measurements by means of 4.5'' sensors by Spectra Symbol and determined a cubic fitting curve for forces within the 0–4 N range.

In general, RFSs are used at room temperature conditions, but for particular applications when the temperature can meaningfully vary, the RFS resistance R behavior with respect to the temperature T ('R versus T' curve) must be taken into account. The work in [25] addressed this problem, working with indigenously made silicon-overlaminated RFSs, based on carbon electrically conductive ink engineered onto polyester plastic sheets. The authors found a 30% resistance variation with temperature increasing from 40 °F (4.4 °C) to 120 °F (48.9 °C).

3.1.3. Step response decay and measurement uncertainty

From a physical point of view, attention has to be focused on the mechanical properties of RFSs, in particular on their weight, durability/robustness and environmental stability.

Manufacturers claim high robustness in their RFSs, stating that they are capable of flexing and extending millions of times. No proofs of such lifetime are provided, but RFSs were nevertheless adopted to measure liquid flow rates continuously for three weeks without maintenance [92], and were used as goniometer tracers for index finger, wrist, and elbow on 21 subjects over the course of six months without requiring replacements [91].

Base RFSs demonstrated aquatic environmental stability when adopted in a test flume, with the intent to map out the flux of dissolved chemicals and solids [93] typically under at most 3 m of water [92]. Stability was not recorded with respect to temperature variations to which RFSs are rather extremely sensitive [25], so we should assume they can be adopted as temperature sensors too.

However, when necessary, environmental stability enhancements can be realized by means of extra protection, obtained by over-laminating or over-molding with a supplementary layer added to the standard sensor coating. This extra protection can also help to reduce the sensitivity to mechanical scratches that can meaningfully affect the electrical properties. However, any additive layer can reduce, even substantially, the sensor sensitivity.

A key question is how the mechanical properties of the substrate influence the electrical behavior of the RFSs. This is to allow some predictability in performances. This aspect has been treated in just one work [87], which demonstrated that the more the substrate is bent the more the resistance of the conductive part increases, in a linear way. Any nonlinearity in the electrical behavior (as treated in section 3.1) is only dependent on cracks in the resistive element during mechanical stress.

From a biochemical point of view, it is interesting to use the properties of RFSs, adopting them under human or animal skin, realizing e.g. a pervasive monitoring implantable device for wireless capsule devices [94] or 'CardioMEMS', which was recently developed [95]. In fact, permanent or semipermanent insertion can provide a better quality of life for those who have experienced severe physical trauma, when it is mandatory to record a stream of data from particular joints so as to continuously monitor reduced motor activities, and/or to drive electromechanical prosthesis. Only one work has addressed this problem, performing in vitro analysis, and it was demonstrated that only specific RFSs (polyimide-encapsulated ones) can be potentially used for this purpose because they showed low pyrogenic effects, with no relevant cytotoxicity, and displayed a low proinflammatory profile [96].

Measurement of the physical activities of the upper limb with resistive flex sensors involves mostly the hand, which is by far the most investigated part of the entire body. The hand is a masterpiece of highly articulated mechanical joints so that measuring its motion is a very challenging matter. Hand data have been generally collected by means of a so called sensory glove (also known as an instrumented, goniometric or data glove), which is a nonobtrusive glove equipped with sensors. The key aspects of such a glove are the sensor placing, the glove material, and the number of DOFs to be investigated, considered in the following.

The most commonly adopted kinematic hand model considers 27 DOFs for the hand [140]. Assuming the joints to be spherical, four DOFs are for the index, middle, ring and little fingers: the metacarpo-phalangeal (MCP), the proximal-interphalangeal (PIP) and the distal-interphalangeal (DIP) joints, capable of three flex/extension DOFs and one abdu/adduction DOF. Five DOFs are for the thumb: the metacarpo-phalangeal (MCP) and the inter-phalangeal (IP) joints, capable of three flex/extension and two abdu/adduction DOFs. Six DOFs are for the wrist (rotational/translational capabilities) (figure 11). Therefore, theoretically, 27 sensors, one for each DOF, are necessary.

Zoom In Zoom Out Reset image size

To perform finger flex/extension tracking, RFSs are singularly placed upon the dorsal aspect of the finger joints following their movements. Each sensor is embedded in a sheath, sleeve or pocket, made of nylon [141] or Lycra® [88], so as to be free to follow the joint flexion without side drifting, and generally sewn on the dorsal aspect of a glove (one exception can be found in [131], where sensors were applied in the internal part of the glove by means of light-strips). In contrast to flex/extension, measurement of abdu/adduction capabilities of fingers has been a challenging task, due to the difficulty in effectively locating the RFSs. Up to now, two proposals have been considered: 1) laying each sensor along its narrow (few μm) side between two adjacent fingers passing on top of the hand web spaces [131] (figure 12(a)); 2) fixing both ends of each sensor on top of the dorsal part of two adjacent fingers so as to realize a half circle which varies its radius with the splaying of the fingers (figure 12(b)) [88, 142]. The first proposal is mechanically hard to realize, the second proposal obstructs object grabbing actions, and neither proposal has ever been investigated from the point of view of measurement reliability.

Zoom In Zoom Out Reset image size

Evidence-based practice shows that the glove must be lightweight, minimally invasive, stretchable, and comfortable. A crucial parameter is the tightness so as to guarantee that the sensors are tightly pressed over the fingers to capture the finest flexion angular change [62]. Different materials have been investigated, mostly Lycra®-based, such as a pliable [78] or stretchable grade of Lycra® [88], a mix of 93% Lycra®/7% Nylon blend [71], a Lycra®/cotton combination (spandex) [79, 141], a polyamide/Lycra® combination [76]. Others are neoprene [102], stretchable velvet [143], silk [144], a lightweight thin leather golf glove [145] and even a sheep skin [131] or a material used by bikers (by Fox Head Inc., Irvine, CA) [110].

However, RFSs can be used without a supporting glove, just inserted into an elastic textile tube anchored using Velcro and thick foam rings to keep them next to the finger [102], or inserted into a thin sleeve attached to the back of the finger with a double-sided toupee tape adhesive [71].

Actually, lateral movements of the thumb have usually been approximately considered only related to the trapeziometacarpal joint, neglecting the ones due to the thumb metacarpophalangeal joint. Without considering the wrist, the model reduces to 20 DOFs. This number can be further decreased by taking into account intra-finger, angle-range and one-axis constraint equations among joints [146], since fingers cannot make any arbitrary movements. Additional reductions are possible focusing only on a limited set of movements not involving all the hand capabilities. Therefore, in the literature sensory gloves have been set up with different sensor numbers.

Just one RFS was used to track a coarse middle finger movement made by the patient to report a perceived pain sensation when subjected to a Digitest^TM electric pulp tester [147]. One RFS aided in providing a closed loop control of a limb [148]. One RFS was used between thumb and index fingers to monitor hand closing [149]. One RFS was placed on the dorsal aspect of the forefinger of a glove, utilized as a smart attender calling system for patients capable of activating a voice module flexing the finger so to alert caregivers [150].

Only two RFSs were inserted into a sensory glove for monitoring hand rehabilitation [151] or for driving a prosthetic hand [152].

A therapy system developed for upper arm rehabilitation in children was based on three RFSs on the thumb, index, and middle fingers [102], and three RFSs were sufficient to discriminate grasp movements in [153].

Four RFSs were utilized for the first commercialized glove, 'Power Glove' by Mattel Intellivision, dedicated to the Nintendo game console, and four joint angles were measured for microsurgical procedure analysis [145]. In [110], a glove for bikers was equipped with four RFSs to capture the flex/extension of three digits and the addu/abduction movements of the thumb. Again, four RFSs were adopted in support of the assessment of a finger joint angle interpretation method [154], as well as to develop an electrical stimulation grasp-assistive glove to be used in robotic stroke therapy [141].

Five RFSs were necessary to realize an updated version of the 'Power Glove', named the 'P5 Glove' by Essential Reality Inc., which enhanced the precision of its predecessor, reaching 0.5° in angle, and five RFSs integrated into a manifold system designed for capturing gestures [155], as well as for a finger posture wearable monitor [71].

Six RFSs were integrated into a system to support impaired individuals by allowing them a limited ambulatory function [131], and they were also used to monitor finger flexion and palmar and dorsal flexion of the wrist for neurorehabilitation purposes [76].

Seven RFSs were utilized to design an unrestricted upper-extremity rehabilitation system [79].

Nine RFSs equipped the so-called 'SignSpeak Glove', in addition to 11 contact sensors and one accelerometer, to realize a sign-to-speech conversion [156].

Ten RFSs aided hand gesture recognition projects [157, 158], equipped the so-called 'BabyGlove' to assess hand motion control development in infants [68], and ten RFSs were necessary to design a mastering glove to control robots through hand gestures [159]. In [144], a study tested the astronauts' hand dexterity and mobility during extravehicular activity while dressed in different kind of gloves, with five metacarpophalangeal and five proximal interphalangeal joints recorded by means of ten RFSs. The 'PneuGlove' was a pneumatically actuated glove used to assist an impaired hand to control a virtual hand playing a set of keys [160]. Such a sensory glove provided measurements of each digit by means of ten RFSs, and pneumatic actuation by means of servovalves.

Fourteen RFSs were integrated into a sensory glove system for the assessment of human finger movements in neurophysiological settings [88], and were part of novel array structures in a goniometric glove [58].

Complete sensory gloves with twenty RFSs were used for clinical hand assessment [78], and were utilized in a robotic hand system working under a master–slave configuration [109]. Twenty RFSs were part of a complex sensory glove including 16 tri-axial accelerometers and 11 force sensors too, for arthritis rehabilitation applications [59].

Finally, the 'CyberGlove', by Virtual Technology (San Jose, CA, USA), is a commercial product equipped with 18 or 22 (the number depending on the version, the 22 DOFs including palm arch measures) proprietary sensors, presumably carbon-ink based, furnishing an angle resolution of less than 1°.

Few works report the utilization of a glove equipped with sensors dedicated to the wrist motion capabilities [76, 161, 162], and only one concerns a sensory glove used to measure hand capabilities of small primates rather than humans [104].

Table 2 illustrates a full picture of RFSs embedded in sensory gloves, reporting the number of sensors used, the kind of measurements which have been performed, and the manufacturer companies.

Sensory gloves are valuable when they guarantee repeatability and reliability of measurements. The repeatability is defined both in terms of range (R, the highest-lowest angle value difference among repeated measures) and standard deviation (SD). The reliability, that is, the degree of consistency between successive measures of the same joint movement, is based on the intraclass correlation coefficient (ICC) value. The best values reported for noncommercial gloves have been R = 3.25°, SD = 1.07° [58] and ICC ≅ 1 [118]. It is interesting to compare these results with the usually adopted method of measurements based on visual estimation and mechanical goniometry. In particular, 40 therapists obtained, with visual estimation, an average of R = 80°, SD = 16.85°, and with a mechanical goniometer R = 7.5°, SD = 1.4° [125]; four student volunteers used a professional mechanical goniometer to measure the ROM of 27 young adults, obtaining an ICC = 0.96 at their best [163].

According to the electrical characteristics of RFSs (discussed in section 3.1), a calibration (and eventually normalization) step is necessary to render the measurements unaffected by the differences in user's fingers, and unrelated by time drifts.

RFSs can also be utilized to measure other upper limb capabilities even if few applications have been developed accordingly. In [164] data from an RFS measuring elbow capabilities were compared to the ones measured by a manual goniometer, and in [165] an RFS realized a monitoring system and data logging of an elbow guard. In [166] an RFS, used to sense elbow movements, integrated a motion-based game control system. Reference [167] reports the correlation between the volitional motor signals and the forelimb movements in rats, measured by means of an RFS.

For completeness, the interested reader can find a comprehensive study of the biomechanics of the upper limbs in [168] and a description of the human fingers ROM in [169].

RFSs have been used to measure movements related to the head and torso.

We start by mentioning the monitoring of swallowing-related laryngeal movement. Currently, video-fluorography examination is the most widely-used method for detecting aspiration linked to dysphagia. In [170] a noninvasive method was proposed for assessing swallowing function/dysfunction by means of a flex sensor fixed on the skin surface along the midline of the neck. The results demonstrated accordance between data from the video-fluorography system and data from the RFSs.

Four RFSs were placed around the neck to obtain information about head flexion angle in [113], in order to recognize the head position in a project aimed to classify the eye gaze direction, comparing SVM, neural network and decision tree classifier outputs. Thanks to the introduction of the RFSs, at 20° head flexion the authors meaningfully enhanced the accuracy of the classification from 50% to 95%.

A nontrivial matter is the reconstruction of the shape and measurement of the movements of the torso, because of its complex form with capabilities of multiplanar axis motions. The human torso locally changes in curvature radius with movements, which is different to a joint which always presents the same curvature radius. In [171] the aspect was addressed (figure 13), and in [73] RFSs have been demonstrated to be effective for this kind of measurement too.

Zoom In Zoom Out Reset image size

A thorax reconstruction and the record of breathing movements were treated with a series of RFSs placed around the chest of a volunteer [19]. A recent work implemented a network of RFSs to acquire patient-specific thorax boundary information, demonstrating a shape error of less than 6% with respect to the total perimeter [18]. A system made of several RFSs attached to a wearable belt was used to follow the deformation angles due to the foetal movement during the last weeks of pregnancy [130]. A gyroscope and an RFS were placed from the torso to the upper arm in order to measure flex/extension of the shoulder in [149].

However, due to some response decays of RFSs when held in the flex position (see section 3.1), and according to a study that evaluated the performances of different RFSs [83], plastic optical fiber sensors appear to be more suitable for wearable trunk posture sensing, even if the latter is performed with a good degree of accuracy [172]. Other measurement systems were preferred to capture the contour/3D shape of the back for patients suffering from musculoskeletal deformities that require customized medical seating for accommodating orthopaedic deformities. In fact, in a comparative analysis of design concepts, the solution using RFSs was considered as 'unsatisfactory', because of the difficulties in implementation [17].

Different segments of the human lower limb have been measured by means of RFSs.

We start by discussing the knee, to which a great deal of attention has been dedicated in order to find the most effective system of measurement. In [164], data from a single RFS were compared to the ones coming from a goniometer. The RFS was aligned in the center of the knee joint, with a plastic tube working as a guide to allow the sensor to move accordingly with knee flexion. In [105] an RFS-based goniometer was two-point fixed to two transverse bars, with respect to the limb, for sensing the knee joint movements. In [166] a band provided with an RFS was worn on the user's knee to realize a motion-based game control system. The project in [121] recognized sitting, standing, and walking activities, by means of a resizable RFS band worn on the user's knee, obtaining an accuracy of 81%. In [80] the inner angle of a knee joint was measured by means of two sensors mounted on a supportive cloth. Data indicated slightly different behaviors of the two sensors, so that an extended Kalman filter was adopted to merge the different results. A comparison with data coming from a mechanical goniometer provided a correlation of 0.98, and an error of 6.92°. Anyway, the accuracy obtained with RFSs was lower than the one obtainable by means of inertial measurement units (IMUs), which performed with a range of average error within 0.08°–3.06° [80]. Another limitation was highlighted since the RFS response delay was higher with respect to the rotation times of a knee of a running man at the maximum speed of 10 m s⁻¹ [67]. In any case, the position of the sensor utilized was found to influence the accuracy of the measurement [121].

In order to detect the movements of the leg, RFSs have been adopted not only to follow joint flexion, but also to detect muscle shape changes, as in [138] where the semitendineous muscle was investigated for its large shape changes with walking.

Only one work has been devoted to ankle motion monitoring with RFSs [126], but several works have investigated foot mechanical capabilities integrating sensors in footwear. In particular, two back-to-back RFSs, inserted into an insole underneath the arch of the foot, provided information about flex/extension foot movements [63]. The back-to-back solution of two RFSs was also adopted in [65], where the sensors were differently located, to evaluate flexion at the metatarsophalangeal joint, plantar and dorsiflexion. An RFS was mounted on the insole position of a shoe for plantar flexion evaluation during walking, underlining differences between normal and abnormal walking styles based on pattern recognition algorithms [120]. An RFS was used to measure the movements of a middle toe used as a signal control of a prosthesis in [138]. The signal, of about 150 mV in amplitude, was very clear for stationary settings, but much larger for voluntary flexion than from walking tasks.

Proximal or remote control, in a system with RFSs capable of providing information about human motor activities, has enhanced many human activities.

As such, different human–computer interactions (HCIs) can be realized. We start by mentioning the simplest application that, by means of a sensory glove equipped with RFSs, captures users' hand movements to activate corresponding operations during a power-point presentation [119], or to replace mouse functions [101]. In [62], a sensory glove is developed for a project aimed at controlling a robotic manipulator. In [173] some keyboard and mouse functions were replaced by a sensory glove capable of clicking, dragging, rotating, and pointing to an ideal position on the screen, thanks to data collected by RFSs and further classified by means of a k-nearest neighbours (k-NN) classifier.

Other applications have realized a motion control of a motorized object with four wheels [174] or, similarly, of a wheelchair [107, 175, 176] moved on the basis of fine finger movements rather than hand joystick maneuvering.

Enhanced applications were related to the one-to-one mapping of human to robot motion with functional anthropomorphism. These implementations can be realized with different technologies indeed, reviewed in [177] and mathematically formalized in [178, 179]. However, the simplest human motion capture systems, intended to be the first step towards realizing these dexterous manipulations, were built on the basis of five RFSs, combined with potentiometers and accelerometers in [180], and 20 RFSs in [181]. In [182], flex/extension data from a sensory glove, together with 3D-positional data from the Microsoft Kinect™, enhanced fine-point control of the end effector of a simple robotic claw, in a more complex project aimed to help persons with disabilities. In [109] a human operator, by means of a sensory glove, tele-operated a robotic hand system working under a master–slave configuration. The anthropomorphic robotic hand had fingers capable of performing motions including flex/extension, abdu/adduction and circumduction. The angle ranges for both human and robotic hands were considered 'approximately equivalent', but no further explanation was to define this 'equivalence' and no meaningful data comparisons were furnished. Conversely, a work dedicated to space extravehicular activity and based on 14 RFSs [183] argued the feasibility of telemanipulations, demonstrating that state-of-the art sensory gloves perform with errors of 4.6° on average, increasing to 6.5° when replicated by off-the-shelf anthropomorphic robotic hands. A paper was dedicated to a slave humanoid robot having the capability of sensing the environment and responding to human commands [184]. It was teleoperated by a wearable system made by several types of sensors, including RFSs used for measuring the bending angles of elbows and fingers. The work in [185] considered the control of robotic hands in a closed-loop master–slave scheme, where the robotic hand was the slave reacting to orders coming from the sensors integrated into the glove of the human operator as a master.

A completely different telecontrol was proposed in [186], where specific movements performed by patients were measured by means of accelerometers and RFSs fixed on their fingers, and the relative signals sent to a base station unit which provided consistent audio messages to the medical staff.

RFSs have been successfully adopted in virtual and/or augmented environment systems, to realize simple visual feedbacks, virtual manipulations, complex systems to support human motor rehabilitation, and to create more immersive digital games.

In particular, RFS-based sensory gloves were used to measure the hand dexterity of both expert and novice surgeons, in order to discriminate their ability and to provide an objective surgery evaluation method [187]. The hand movements were replicated by two different virtual avatars on a computer screen, so as to provide a visual feedback to the novice about the differences of his/her gestures and the gestures of the expert (figure 14).

Zoom In Zoom Out Reset image size

A certain degree of interactivity was obtained in virtual manipulation systems realized on the basis of RFSs. In [188] a 3D virtual osteotomy was implemented, with a surgeon driving a scalpel virtually held in a digital gloved hand equipped with RFSs. After that, real-time computer graphics CAD/CAM applications interpreted the flexion angles of the finger joints, measured by means of the 'CyberGlove' (see section 5.1), for novel multidimensional interaction devices [189]. Again, the 'CyberGlove' supported full hand posture estimation for a project dedicated to the realization of a real-time interaction with a virtual rigid objects system [190]. In [191] virtual prototypes were manipulated and modified on the basis of the hand gesture inputs, measured by sensory gloves, for a gesture-based augmented reality design. Jung and colleagues recognized that the sensory glove plays a key role in the so-called 'egocentric' virtual interaction techniques [192], and Ma and colleagues proposed the application of RFSs and sensory gloves in 3D animation for forensic visualization [193].

A virtual ambience was created to motivate children during pediatric motor rehabilitation, by means of a system based on RFSs for measuring thumb, index and middle finger movements, converted into virtual object manipulations on a screen [102]. Similarly, a low-cost viable hand rehabilitation system was developed within an augmented reality system [13], and a rehabilitation game supported the upper extremity movements recorded by means of RFSs [194].

An RFS-equipped band worn by the user on his/her elbow and knee, and an RFS-equipped belt on his/her waist, enhanced the motion control for a ludic virtual ambience [166], while the same kind of sensors were adopted to record movements of a user standing on a wobble board, in order to provide commands useful to control a digital game [195]. In [160], subjects with chronic hemiparesis exhibiting mild to moderate hand impairment were helped to play a virtual keyboard by means of a sensory pneumatically actuated glove (with RFSs and servovalves), named 'PneuGlove'. These subjects demonstrated statistically significant improvement in motor task performance after 3 × 18 h-long sessions per week for six weeks.

For completeness, the interested reader can find a review of interaction techniques in virtual environments in [192].

Humans use their hands with a high degree of dexterity, to grasp and manipulate objects, to execute daily tasks, or even to perform body-language-based nonverbal communications. Currently, this dexterity can be measured by means of webcam-based tracking contours systems. For examples, smart TVs allow us to change channel or to vary the volume with a motion of the hand; some tablets allow music browsing by the user waving their hand. RFSs might open new commercial applications, in particular when integrated into systems useful for motor activity recognition and evaluation, for gesture assistance and interpretation, and for character set and sign language interpretation.

In [76] an RFS-based glove was developed to capture the details of grasping an object, which was useful for clinical practice and rehabilitation settings. The glove reliability was determined by ICC values ranging between 0.84–0.92 and the accuracy error was ±5°. Another work estimated the movements of three fingers during the grasping of an object, including principle component analysis (PCA) to minimize the dimensionality of the dataset [99].

A very challenging application is the measurement and analysis of hand coordination in surgical and microsurgical practice evaluation. A system based on a sensory glove and classification algorithms might score a surgeon's manual skillfulness better than the judgment of experts, which can be biased or prone to human error. In [196] the author developed a glove equipped with six RFSs and nine inertial measurement units, to capture hand gestures in a neurological simulation environment. In [145] a sensory glove was used to measure the angular flex/extension movements of the interphalangeal joints and the grasping pressure of the hand, to analyze microsurgical hand coordination patterns. Similarly, in [197, 198], a glove utilizing fourteen RFSs was used to determine the typology of surgical gesture by assigning it to a predefined class.

Fingers and whole hand movements can be recognized with some differences by means of vision-based analysis, as reviewed in [199], or by means of data acquired with RFSs, as in [200], to realize a gesture-based interface to control household equipment.

RFSs can be used in electronic systems capable of interpreting a character set, so as to help the communication of speechless users. The characters of the American sign language alphabet were acquired by means of a glove with 21 sensors, out of which nine were RFS, and PCA was used to classify the real time input data for feature extraction, with an accuracy of 95% [156]. In [201] the accuracy was equal to 95.13% for the recognition of 42 static hand gestures by means of a neural network classification algorithm receiving a combination of data from a camera-based system and a commercial 5DT data glove (by Fifth Dimension Technologies, Irvine, CA, USA) equipped with 14 RFSs.

The conversion of the Spanish sign alphabet into voice and writing representation was realized by means of a glove equipped with ten RFSs, implementing the discrimination of signs by means of four different multilayer perceptron neural networks, with a classification accuracy of 97.39% at best [202]. In [203] a glove with five RFSs was developed to map bend measurements to the American sign language alphabet, with the user maintaining a static posture for two seconds to improve recognition. A 3D accelerometer was added to a glove with five RFSs in order to add the motion and orientation of the hand to the finger bending measurements and the system was validated with data from nine speech-impaired subjects [200]. A supervised clustering classifier yielded an accuracy of 91.53% for subject-specific models, and an averaged accuracy of 95.34% across subjects. Another work proposed a portable system consisting of an RFSs-based glove able to discriminate up to 243 messages [204]. A similar work was dedicated to the conversion of finger-spelled words into a language processing module, phonetically transcribing each phoneme of the related word of the Malay language [205]. In [206] a real-time system was realized that was capable of recognizing static hand gestures corresponding to the Italian manual alphabet of sign language. Each gesture was mapped onto the corresponding alphabet letter by comparing three different classifiers, namely support vector machines, Mahalanobis distances and Euclidean distances, obtaining an accuracy of >90%. A complete set of 26 characters were classified, with an accuracy of 95%, by means of a glove equipped with fourteen RFSs and a neural network classifier, in [207]. A set of 31 static gestures were recognized, with a classification accuracy ranging between 77.42% and 99.61%, from a South African sign language dataset, adopting the Levenberg–Marquardt training algorithm and neural network classificatory [158]. A reduced set of six words was recognized using the movements of only two fingers and fuzzy logic as a classifier, in [208]. A recent work mapped finger bend data, coming from ten RFSs, into a character set, by implementing a minimum mean square error-based classifier, with a text-to-speech conversion realized on an Android phone [209]. The work proposed in [210] developed a pair of 'gesture vocalize gloves', equipped with five RFSs along the length of each finger and the thumb. Data from sensors were compared to a library with a list of known postures, to assign a particular character to a particular gesture.

More complex is the understanding of dynamic hand and finger configurations, with respect to the aforementioned static ones, to realize conversions of sign languages into speech, and not only into characters. The work in [157] provided the first approach with the classification of eight dynamic gestures, in addition to 47 static ones, of the Taiwanese sign language. The idea was to reduce the problem from 3D to 2D patterns, projecting the sequence of a hand gesture into a 2D trajectory, and then turning a gesture recognition problem into a hand-written characters classification. The work realized in [211] classified a set of 2000 dynamically-executed hand gestures of the Italian sign language, on the basis of a fifteen RFSs-based glove and on wavelet analysis (WA) plus support vector machine (SVM). The accuracy was 86.4%, incremented to 98.8% when adopting artificial pauses between gestures.

In [212] a pattern matching technique was used to implement sign language tutoring rather than a gesture sign interpretation. Animated 3D models from four teachers, performing 65 common gesture signs, were replicated by ten trainees obtaining an average accuracy of 93.8%. The use of a sensory glove, incorporating ten RFSs and ten tactile sensors, was investigated in [213] to assist children learning the Auslan sign language.

The study reported in [116] empowered the user to figure out for themselves the information they want to capture by means of a set of sensors, and utilize that information in the developing of new applications for personal informatics. The study found the most effective sensors to capture helpful data in the daily life context to be the simple light sensor, the sound sensor, the RFID reader and the RFS. In [214], mute people were assisted to vocally express themselves by means of a system that, based on the amount of bending exerted on an RFS, delivered different stored messages to a speaker.

Personal robots, used for fundamental assistance, have to provide smooth and natural interaction with their human partners. With this aim, measurements made on humans, i.e. head, arms, hand movements, as well as physiological parameters, can provide basic information to determine the robot response. In particular, in [215] the movements of the fingers were addressed by means of a light sensory glove in a fifteen DOFs configuration using fifteen commercial RFSs which performed with an error of <2° in measuring flex/extension activity.

In recent years, the growth in the number of automobiles on our roads has vastly increased the number of injuries sustained by car occupants and pedestrians. According to a recent report [216], there are nearly 1.3 million deaths on the world's roads, and 20 to 50 million injuries, each year, with an expectation of those numbers doubling by 2030 unless counteractions are taken. Therefore, brand new safety measures could help in reducing the number of road injuries and traffic-related deaths. With this is mind, RFSs have been considered to be of some interest for the automotive industry, in view of car occupant safety and traffic safety, as witnessed by some research works and patents.

7.1.1. Car occupant safety

Regarding car occupant safety, in 1999 a patent proposed a seat belt tension measurement apparatus capable of inhibiting the air bag deployment upon the detection of a predetermined seat belt tension, measured by means of an RFS [217]. Also, in [218] a combination of sensors was proposed, including flex ones, to discriminate passengers from carried objects so as to correctly switch on/off a seatbelt reminder beeper. The so-called 'smart car seat' project aimed to prevent a child from being left alone in a vehicle during extreme temperatures events [219]. It consisted of sensors (including flex ones) and electronic circuitry used to alert parents via a buzzer when a motion was detected and the temperature reached a maximum pre-defined value. In [220] RFSs were adopted to determine the mat pressure profiling of 'smart airbags'. Current airbag systems are designed considering only the 50th percentile of male seat occupants, therefore the work regarded the suppression of airbag deployment for small occupants in some occurrences, and was established on the basis of the responses of flex sensors. In addition, according to the National Highway Transportation Safety Administration (NHTSA) [221], the adoption of the RFS mat could avoid the speed and inflation force of passenger-side air bags, which are responsible for the deaths of a certain number of children and short adults. Flexpoint Inc. reports interest in seat mats from Germany, France, Spain and some European corporations, as well as important companies such as General Motors, Ford, Chrysler, TRW, Lear, Takata, Magna, Allied Signal, Autoliv, Johnson Controls, Siemens, and Delphi [221]. Figure 15 shows an example of an RFS mat.

Zoom In Zoom Out Reset image size

7.1.2. Traffic safety

In recent years, an increasing number of projects about robots have involved RFSs, generally applied to detect obstacles and sense positions, to take part in closed-loop control systems, or to develop the abilities of humanoid robots.

Some animals, such as rodents, rats and cats, use a vibrissal system, based on stiff hairs located on their face, as a tactile organ to sense the environment. Starting with the work of Russell [226], several systems inspired by the animal vibrissal system have been designed to be included in robot machines, some of which utilize RFSs in place of an animal's stiff hairs. In [66, 227] the so-called 'Darwin IX' robot utilized RFSs as whisker arrays so that the robot could sense and follow wall patterns. In [228] a robot was developed with a biologically-inspired legged locomotion system, named 'Robot V', where RFSs were rigidly attached on either side of some joints to implement a part of a position feedback mechanism. RFSs overcame a problem of the previously realized 'Robot IV' that based the angular readings on rotary potentiometers which needed frequent recalibrations. The 'Dynarobin' project developed a quadruped robot with elastic spiral-shaped feet [229]. The contact angle adaptation algorithm of each foot was based on the contact time feedback measured by a flex sensor, which also served as a rudimentary approximation of the ground stiffness coefficient. The 'Map Bot' is a robot capable of mapping out a floor and then using that map to traverse the mapped area [230]. This robot was based on an ultrasound sensor and two RFSs (figure 16).

Zoom In Zoom Out Reset image size

The so called 'Garcia Robot' utilized the same controller that governs the following of a wall by a cockroach, by means of an antenna that provides tactile feedback, integrated along its length with a series of RFSs that measure the local curvatures of the antenna [231]. Such a simple mechanism successfully allowed a wheeled robot to follow a wall at a speed as high as the one of Nature's fastest terrestrial insects. In [69] a mobile robotic system detected both its contact state and the direction of contact parallel to the movement plane by means of a fuzzy logic-based sensor system, realized on the basis of an array of RFSs. With the perspective of sensing the spatial position, a recent patent proposed a robotic submersible cleaning system, fitted with RFSs to detect underwater inlets and outlets [232]. In [233] RFSs were considered a compact lightweight, cost-effective embedded solution to implement the sensory feedback in an artificial finger for prosthetic and wearable robotic hand applications. An RFS was mounted within the upper inner layer of the artificial proximal interphalangeal finger joint to act as an proprioceptive sensor, providing positional feedbacks. An RFS was proposed in a patent aimed at determining the (applied external or actuating internal) force exerted on a robotic arm [234].

In recent years, several robots with the appearances of animals have been realized. In [72] a biomimetic fish was realized with the idea of implementing a closed-loop system to control the movement of the tail according to data motion acquired by means of an RFS.

Some new abilities of humanoid robots can be achieved on the basis of human imitation, as in [235] where a wearable interface was developed to teach humanoid robots by virtue of user demonstration. The wearable interface was equipped with some low-cost sensors, such as inclinometers, magnetic and RFSs, as motion capture devices. In particular, RFSs were part of a sensory glove to measure the bending degrees of human fingers mapped onto robotic ones. The project was also enriched with a stochastic model used to generate motions, such as heart drawing and boxing, avoiding the monotonous repetition of demonstrated gestures. Similarly, a later application intended to demonstrate a general method to map hand motions to a robotic arm, by means of a sensory glove equipped with two RFSs and a three-axis accelerometer [236]. Finally, a recent work [237] investigated the correlation, by the use of the regression method, between the angle measured by a reference motion capture system (MOCAP) and the signal produced by an RFS attached to a robotic hand, thus demonstrating the validity of the outcomes from the RFS.

The work in [238] developed three wheeled mobile robots for a mechatronics education system. The motion of each robot was based on the change in environmental conditions, determined by means of six position sensitive detector (PSD) sensors and two RFSs.

7.3.1. Orthosis

7.3.2. Prosthesis

In the area of prostheses, one of the first prosthetic hands integrating touch and flex sensors inside the robotic fingers was developed at the Forschungszentrum Karlsruhe research center [137]. That hand was capable of flex/extension of the fingers in less than 100 ms, resulting in an open/close hand movement at >5 Hz, faster than the world's current most advanced prosthetic hand (by Bebionic, San Antonio, Texas, with a 0.5 s minimum open/close time), with a very close approximation of the ROM of a human hand. These RFS performances rates are in contrast with the ones treated in [89], where the RFSs demonstrate a latency time when used in recording movements faster than 2 Hz. For the prosthetic hand project a key request was lightness, so that each finger weighted less than 20 grams and the lightweight of RFS was strategic to the prosthetic hand project.

Another work [108] proposed to identify the movements of an artificial arm, from flexion and extension, pronation and supination, by means of RFSs rather than the commonly-adopted surface electrodes. Experimental results demonstrated a maximum error of 17° in measuring flexion and extension angles, and 10° in pronation and supination.

In order to perform a reaching or a grasping movement, humans rely on different sensory channels, such as tactile, vision, acoustic, proprioceptive, and vestibular channels. Similarly, to move towards natural movements with a prosthetic hand, he/she needs feed-forward (intentions) and feed-back (time-space position) information, but amputees have lost part of their proprioception. Therefore, in [152] the use of audio feedback was investigated to improve the learning of new perceptions, without relying on visual input, and the user's finger spatio-temporal information was captured by two RFSs.

A project concerning the improvement of the amputee's control over their prosthetic limb was presented in [138], with the aim of discovering suitable methods for identifying user intent and moving the prosthetic hand accordingly. It was fundamental to measure flexion capabilities, so some efforts was devoted to the use of RFSs, even though the ultimate choice relied on mechanomyography (MMG) sensors.

Reference [241] reports the realization of a multigrasp myoelectric controller to drive a virtual transradial prosthesis through surface EMG electrodes. For comparison purposes, the same virtual prosthesis was moved using data provided by a sensory glove, equipped with four RFSs, used to capture the flexion/extension of the thumb, index and middle fingers, and the thumb opposition. In a subsequent work [110], the authors assessed the efficacy of the system with five non-amputee subjects as users in a series of experimental trials, still using the sensory glove (figure 17).

Zoom In Zoom Out Reset image size

The last work concerning prosthetic control including the use of RFSs was developed for amputees who are missing some of their digits [242]. The authors measured the movements of the remaining healthy digits by means of RFSs mounted onto a glove. The flex sensor readings were compared to a set of training data to determine the user's intent, so that bionic digits were moved accordingly. Mean success rates were 96.5% with one amputated digit and 92.5% with two amputated digits.

RFSs have found different applications for gesture and motion-based wearable controllers, with the aim of realizing music-related tools for musicians [75, 111, 243–247], dancers [63, 112, 248], vocalists [143, 249], or for laypersons [250] and impaired people [251].

We start by considering a project using a unique RFS placed on the dorsal aspect of the index finger, so that its flex/extension movements actuate the simple increase/decrease in volume of an audio system [247]. The project used the three DOFs of one finger at one time, considering them to be correlated so that only one RFS was adopted.

The standard method to create music requires the player to acquire expertise, spending a huge amount of time in learning specific gestures and movements to play specific instrument. A different approach comes from the possibility of transforming common, nonspecific gestures into sound creation and/or production. In [243], the author experienced the use of a self-made flex sensor-equipped glove in a system named 'sign-drum', which was able to convert a predefined set of hand signs into different percussive sounds. The analog values captured by 19 RFSs lead to ten hand shape representations processed by a back-propagation neural network.

In [63], 'expressive footwear' referred to a project which furnished the dancer with a new tool. This was obtained by means of a sensory shoe connected to a base station and a sound synthesizer. The overall system gathered data from the physical parameters of the dancer's feet and transformed them into music in real time. A strict requirement for the project was unobtrusiveness, so the sensors were selected accordingly, in particular including force sensing resistors, accelerometers, electronic compasses, and RFSs. The latter furnished information about foot flexures, used to control the octave of the melody notes. Overall gestures were recognized by applying hidden Markov models to the acquired data.

A two-foot long flexible rubber tube called the 'Sonic Banana', incorporating four RFSs in a row, realized a MIDI controller and an alternative musical instrument [245]. Bending the tube changed the MIDI control values continuously which created a variety of instrumental sounds.

The project in [111] tracked human physical movements and converted them into audio (and graphic) signals, in accordance with a set of mapping rules and algorithms, with the intent of exploring the controllability, interfacing, and expressivity of different musical instruments. An augmented drum prototype, for example, was equipped with RFSs embedded in the drum brushes, in order to enable additional controls and manipulations of real-time multimedia outputs (e.g., MIDI data or graphic/video projection), by means of the information gathered by tracking the bending angle of the brushes during the performance.

A five flex sensor-based sensory glove was used to realize an assistive instrument aimed at improving the functional capabilities of some individuals with disabilities, allowing them to play notes by moving their fingers [251]. The instrument allowed a useful combination of entertainment and therapy for people with cerebral palsy, for whom playing is more demanding than for non-impaired people.

In [246], the authors proposed a wireless sensory glove by means of which the user can control a real time granular sound synthesis process. The glove integrated three mono-axial accelerometers and 19 RFSs, to measure relative hand movements in space, flex/extension of all joints and lateral movements of the fingers, but it was under-utilized because only 17 physical parameters out of the available 21 were considered. Control and synthesizer units were developed under the Max/MSP environment (www.cycling74.com/products/maxmsp) for a real-time musical composition. In two later works [98, 252], the same authors included a mapping strategy based on an artificial neural network, so that the performer did not influence directly the parameters that rule the behavior of the sound generators, as they were preprocessed by the mapping algorithm.

The proposal in [75] was to realize a wireless device that senses user hand movements and distances between hands in order to play music. This was made on the basis of data coming from a sensory glove worn on the user's right hand and a portable and lightweight box on the left hand. The box was provided with a proximity sensor and fret potentiometer, while the glove was provided with four RFSs on each finger, except for the thumb. The overall system simulates in particular a soprano ukulele, the RFSs representing the strings so that by bending a finger, a string picking was simulated.

In [112], drums and bass melody voices were articulated with the control of an instrumented dance sneaker according to foot, sole, and toe movements. Data sources were two pressure sensors, a piezoceramic transducer, an electric-field-sensing electrode, a 2-axis 2G micromechanical accelerometer, a 3-axis high-G piezoelectric accelerometer, an electromechanical compass, a small vibrating-reed gyroscope, and a pair of back-to-back RFSs, all realizing an array of sensors able to acquire several parameters, updated at 25 Hz, of the foot (figure 18).

Zoom In Zoom Out Reset image size

The commonly adopted commercial RFSs respond monodirectionally, i.e. their resistance varies when bent in one direction (inward bending) rather than in both directions (inward and outward bending), as stated in section 2.5. Therefore, the bidirectional characteristic was obtained by placing two monodirectional RFSs in a back-to-back configuration, a solution rarely adopted. In a later work [64], the authors updated their instrumented dance sneaker, adding more sensors (e.g. integrating a sonar) and increasing the sample frequency to 50 Hz.

In 1991, Laetitia Sonami (www.sonami.net/works/ladys-glove) built her first so-called 'Lady's Glove' for a performance at the Ars Electronica Festival (Linz). Her first realization, the 'Lady's Glove no.1', was made with a pair of rubber kitchen gloves, equipped with five Hall-effect transducers and a magnet. The updated 'Lady's Glove no. 3' was a golden-Lycra® glove containing RFSs sewn along the fingers and wrist. Each RFS was tapped at its center with its two ends free to move, so that each sensor supplied two streams of data. This is an original adoption of RFSs, which has still not been repeated.

In [250], musicians and laypersons were able to control musical expressions (pitch, rhythm and volume) of a melody, by wearing the commercial sensory glove named 'CyberGlove' (see section 5.1), and performing a designed set of hand motions and gestures. Minimal training was necessary for either musicians or laypersons.

Different to all the other works, in [248] the authors intended to measure dancers' interactions with other dancers, rather than the classical measure of the movements of an individual performer (angle of elbow or knee or ankle bend). They located the sensors where potential contact/interaction with other performers was possible. Touch was measured by force sensing resistors at the palm, an RFS at the wrist and an accelerometer on top of the hand. Information coming from the sensors was converted into audio out to speakers and video out to a projector.

In [143], the author furnished a vocalist with the possibility of manipulating in real-time his/her sung notes by means of a worn glove-shaped musical controller, named the 'vocal augmentation and manipulation prosthesis' (VAMP), equipped with an accelerometer, a pressure sensor and two RFSs. The latter measured the bend in the elbow and the forward-backward bending in the wrist. The manipulation of the notes occurred according to a pre-defined gestural vocabulary. Similar goals were acheived in [249], with a singer who processed in real-time the relevant parameters of his voice during a performance. This was possible with a sensory glove, for which RFSs were chosen for their ability to sense larger controlled gestures, like the closing of a hand or the bending of a wrist.

Finally, RFSs have been adopted in systems which are useful for biomechanical measurements of hand dexterity during musical performance; for a review article, see [253].

The interested reader can also find a study in [254] regarding the optimal sensor choice for parameter modulations in digital musical instruments. The work established that, considering the overall sensors used in existing digital musical instruments, about 8% of these are RFSs.