Measuring fingerpad deformation during active object manipulation

During active object manipulation, the finger-object interactions give rise to complex fingertip skin deformations. These deformations are in turn encoded by the local tactile afferents and provide rich and behaviorally relevant information to the central nervous system. Most of the work studying the mechanical response of the finger to dynamic loading has been performed under a passive setup, thereby precisely controlling the kinematics or the dynamics of the loading. However, to identify aspects of the deformations that are relevant to online control during object manipulation, it is desirable to measure the skin response in an active setup. To that end, we developed a device that allows us to monitor finger forces, skin deformations, and kinematics during fine manipulation. We describe the device in detail and test it to precisely describe how the fingertip skin in contact with the object deforms during a simple vertical oscillation task. We show that the level of grip force directly influences the fingerpad skin strains and that the strain rates are substantial during active manipulation (norm up to 100%/s). The developed setup will enable us to causally relate sensory information, i.e. skin deformation, to online control, i.e. grip force adjustment, in future studies.


37
Humans are exquisitely skillful at dexterous manipulation of objects with their fingers. This dexterity partly relies 38 on optimally adjusting the grip force (GF), the force exerted normally to the object surface, to the load force (LF) 39 due to the object weight and inertia, and the fingertip-object contact friction (Johansson and Flanagan 2009a). 40 Indeed, an excessive amount of force is undesirable as it leads to excessive expenditure of energy and could 41 potentially crush the object. However, an insufficient grip force will let the object slip from the hands. 42 Accordingly, we usually exert a grip force just above the minimum, thereby continuously varying the GF 43 according to the LF and the level of friction (Cadoret and Smith 1996;Flanagan and Wing 1997;Johansson and 44 Westling 1984;Westling and Johansson 1984). Deformations and vibrations produced in the skin during object 45 manipulation are faithfully encoded by the numerous tactile afferents innervating the hand and fingers. These 46 afferents in turn provide the central nervous system with rich and behaviorally relevant tactile information 47 content (Delhaye et al. 2018(Delhaye et al. , 2021Goodwin and Wheat 2004;Jenmalm et al. 2003;Macefield et al. 1996). This, 48 for instance, allows the central nervous system to adjust the grip force exerted on a manipulated object to the 49 grip conditions (Johansson and Flanagan 2009b). Without tactile afferent feedback, we fail to react to 50 unexpected perturbations in object load or surface friction and to maintain a stable grip force to load force ratio, 51 hence highlighting the critical role played by tactile feedback during object manipulation (Augurelle et al. 2003;52 Nowak et al. 2001;Witney et al. 2004). 53 In vivo biomechanics of the fingerpad has been extensively studied in passive conditions, highlighting the 54 systematic occurrence of partial slips at the periphery of the fingertip-object contact during the tangential 55 loading of the fingerpad (André et al. 2011;Delhaye et al. 2014Delhaye et al. , 2016Tada et al. 2006;Tada and Kanade 2004). 56 Indeed, we have shown that slip is not instantaneous and that the transition from a stable to a slipping contact 57 develops progressively with partial slips initiating at the periphery and progressing towards the center of contact 58 until the point of a full slip. As a consequence of the relative movements between the non-slipping regions and 59 the slipping regions, partial slip is associated with substantial (up to 25%) surface-tangential skin strains (Delhaye 60 et al. 2016). Furthermore, we have shown that those skin strains caused by partial slips are readily encoded by 61 human afferents (Delhaye et al. 2021). Besides, we have also demonstrated that human subjects can 62 consciously perceive incipient slip, before full slip is reached, given that the surface-tangential skin strains are 63 sufficient (Barrea et al. 2018). Taken together, those results have essential implications for object manipulation: 64 indeed, the systematic occurrence of partial slips during tangential loading implies that partial slips will take 65 place during active manipulation. Because partial slips provide a measure of how far from fully slipping the 66 contact is, they might be used by the central nervous system to adjust the GF to the object properties during 67 active manipulation. 68 It remains unclear, however, how much partial slips spread inside the contact area during active object 69 manipulation and how they will be affected by the gripping conditions, including the biomechanical properties 70 of the fingertips. Indeed, for a given object load, the amount of partial slips depends on how much grip force is 71 exerted on the object and also on the frictional properties of the fingertip-object contact (André et al. 2011). 72 To investigate this, we developed a new instrumented device that enables synchronously recording the forces 73 exerted by the fingers together with the fingertip skin deformation at the finger-object contact with a high 74 spatial and temporal resolution during active manipulation. The device was tested in an experiment involving 18 75 subjects who were requested to perform vertical oscillatory movements. We describe the typical strain patterns 76 taking place in the contact in parallel with the kinematic and dynamic parameters of the task. 77

78
Apparatus 79 To characterize the deformations taking place at the contact between the fingerpad and an actively manipulated 80 object, we developed a manipulandum equipped with force sensors as well as an imaging system. This system 81 can capture images of the skin in contact with the object through a transparent and optically flat plate of glass 82 (Figure 1). 83 The device is designed to be handled using a precision grip (Westling and Johansson 1984), i.e. pinched between 84 thumb and index. Its aluminum core supports two force sensors, one on each side of the manipulandum, to 85 measure the forces exerted by each finger (Fig 1B, orange, ATI Mini27 Ti, ATI-IA). On the outer side of each force 86 sensor, an aluminum piece supports a transparent plate of glass, where the object is contacted to be 87 manipulated. The rest of the device, including its top hat and diverse small parts, is 3D printed using PLA to be 88 lightweight and customized at will (Fig 1A-C). Only one finger can be imaged, but the outer plate design is 89 symmetric so that the contacting surfaces are the same for both fingers. Both surfaces can be removed and 90 swapped quickly, allowing the experimenter to test materials of different properties (e.g. friction or texture) 91 during one experiment. An accelerometer (Fig 1B,red,LIS344ALH,STMicroelectronics) is included in the top hat 92 of the device. All signals are acquired using an ADC acquisition system (NI6225, National Instruments). 93 Fingerprints in contact with the glass plate are imaged using a custom optical system based on the principle of 94 frustrated total internal reflection, with a coaxial light source and camera ( Fig 1D). The light source ( Fig 1B,  95 black, LFL-1012-SW2, CCS Inc.) is placed on the opposite side of the imaged finger. The light passes through a 96 half-mirror to illuminate the finger. Part of the light is reflected by the contacting glass and part of it is 97 transmitted. The light transmission index is increased where the fingerprints contact the glass, causing less 98 reflection and therefore dark fingerprints in the captured images ( Fig 1E). The light is then collected by a small 99 and lightweight monochrome camera equipped with a macro lens (Fig 1B,  Note that only a subset of the camera field of view was recorded because of constraints related to the 103 architecture of the device, in particular the size of the area illuminated by the light source. Optical tracking 104 allows us to measure partial slips from fingerprint images and to derive the skin strains (Delhaye et al. 2014(Delhaye et al. , 105 2016. A checkerboard pattern is glued on the glass on the side of the field of view to track very small glass 106 movements relative to the camera caused by the elastic deformation of the device. The squares of the 107 checkerboard, whose dimensions are 0.5 x 0.5 mm, enable us to measure the image resolution (64 pixels/mm) 108 and to verify that the effects of optical distortion and elastic deformation of the manipulandum are minimal. The 109 width of the device, i.e. the distance between both fingers gripping the object, is 50 mm. The total mass of the 110 device is 540 g. 116 and reflected at optical interfaces. In this study, the index finger was imaged, and the thumb was not. E| Typical image obtained with the 117 system. The checkerboard pattern on the right of the image is used to measure very small movements of the glass relative to the camera.

119
Participants 120 Eighteen healthy subjects (9 women, ages 20-34 years) participated in the experiment. Each subject provided 121 written informed consent to the procedures and the study was approved by the local ethics committee at the 122 host institution (Institute of Neuroscience, Université catholique de Louvain, Brussels, Belgium). 123

Experimental procedures 124
For this experiment, the device was held by a system of pulleys and a counter-weight with the same mass as the 125 device, such that the net weight of the whole system was close to zero, but its inertia was twice that of the 126 device alone. The goal was to isolate the forces and deformations related to dynamic interactions (inertial 127 forces) as opposed to static weight. Moreover, an optical distance sensor (DT20-P224B, SICK Sensor Intelligence) 128 was placed above the device and measured its vertical position continuously. 129 Subjects were asked to grab the device with their index and thumb at the center of the glass plates such that the 130 fingerprints were centered in the field of view of the camera (Fig 1C-E). Then, they were instructed to move the 131 device up and down producing an oscillatory movement between two targets spaced by 20 cm for 30 seconds 132 ( Figure 2). The period was set to 1.5 seconds, therefore generating 20 oscillations for each block (Fig 2A). A short 133 tone was played every 750 ms to indicate the instant when they were supposed to reach one of the targets. A 134 higher-pitched tone indicated the end of the block when the subjects were asked to release the object and 135 prepare for the next trial. The same procedure was repeated for 15 blocks. Because we observed that subjects 136 tended to apply an excessive level of grip force, a subset of them (6) were explicitly asked to try to minimize the 137 grip force. 138 Moreover, at the end of the experiment, the subjects were asked to rub their index finger against the glass 139 surface repeatedly using a published procedure (Barrea et al. 2016) to evaluate the coefficient of friction as a 140 function of the normal force. A power function was fit to the data ( = ( ) −1 , where µ is the coefficient of 141 friction). It was used to predict the slip force. 142 The forces, torques, position, and acceleration were acquired at 200 Hz. The images were captured at 100 143 frames per second at a resolution of 1696 by 1248 pixels. 144 145 146 Figure 2: Experimental protocol and image processing. A| Subjects performed vertical oscillations with a peak-to-peak amplitude of 20 147 cm, delimited by visual targets and rhythmed by auditory cues every 0.75 s (at each peak of the movement). B| The device was held in a 148 precision grip. A system of pulleys and a counter-weight completely compensated the weight of the device. C| Image processing pipeline.

149
From the raw images, the contact area is extracted. Feature points are then detected inside this contact area and tracked from frame to 150 frame. A Delaunay triangulation is computed from the first frame and the strain rate tensor of each triangle can be computed for each 151 pair of images.

Data analysis 154
Force, torque, and position signals were low-pass filtered (4 th order Butterworth, zero phase-lag, cutoff 40Hz). 155 The grip force (GF) was obtained by computing the average of the force normal to the surface from the two 156 force sensors (along the z-axis, see Fig 1C). The load force (LF) is the sum of the vertical component of the 157 tangential force from the two force sensors (along the y-axis, see Fig 1C). The object velocity was obtained by 158 numerical differentiation, and the positive velocity peaks were used to split the data into individual oscillations 159 during which different parameters were computed (mean GF, max LF, …). The vertical position of the center of 160 pressure (COP) was obtained using = ( − . 0 )/ , where T is torque along a given axis, F is force and 161 0 is the normal distance between the contact surface and the sensor's surface. It measures the vertical position 162 of the resultant force exerted by the index finger normally to the glass. The slip force (SF) is the minimum 163 normal (grip) force needed to avoid slip. It was obtained from the power function fit for each subject as a 164 function of the tangential force ( = . = ( ) ). The GF and LF modulations were obtained for each 165 oscillation by subtracting the minimum value from the maximum value of the specific oscillation. 166 Image processing 167 We used an image processing pipeline already described in previous work to detect the slipping regions and 168 evaluate the surface skin strains inside the contact area (Delhaye et al. 2014(Delhaye et al. , 2016. It is summarized here (see 169 Fig 2C). First, the gross contact area was extracted from the background using a two-stage procedure. The first 170 stage consists in manually depicting the contact area contour for a small subset of frames (typically 10 to 20 171 frames per subject). In the second stage, features relevant for the segmentation (Sankaran et al. 2017) are 172 computed and fed to a machine learning algorithm (classification tree, fitctree function in Matlab) that is trained 173 on the manually segmented data to extract the contact area and then used to extract the contact area on the 174 whole sequence of frames. 175 Second, for each oscillation, three sets of features having good gradient properties for tracking (Shi and Tomasi 176 1994) were sampled at the first, the middle, and the last frame of each oscillation respectively. Indeed, the 177 contact area varies during the movement due to rolling of the fingertip, i.e. some skin parts are coming into 178 contact and other parts are leaving contact during the active movement. It is therefore essential to re-sample 179 features at different times to cover the contact area with features during the whole oscillation. The minimum 180 spacing between features for feature detection was set to 17 pixels. Then, those features were tracked from 181 frame to frame using a classical algorithm implemented in OpenCV (Bradski 2008;Lucas and Kanade 1981). 182 Depending on the initial frame (first, middle or last), the features were either tracked forward or backward in 183 time (or both). Finally, the 3 sets of features (first, middle and last) were merged, and those overlapping were 184 suppressed. The very small glass movements that occurred due to the compliance of the force sensors and the 185 plates (of the order of a couple of pixels) were monitored by tracking features located on a checkerboard 186 attached to the glass. The velocity of the glass movement was subtracted from the velocity of the fingerprint 187 movement before further analyses. 188 Third, a Delaunay triangulation was performed on the features from their location in the first frame. And the 189 triangle strain rate from frame to frame was computed as described earlier (Delhaye et al. 2016), to yield a 2-by-190 2 strain rate tensor for each triangle and each pair of consecutive frames (Fig 2C). The strain rate norm was 191 obtained by computing the norm of the strain rate tensor for each triangle and each pair of consecutive frames. 192 Moreover, if the triangle's center moved by more than 1/2 of a pixel between 2 consecutive frames, it was 193 considered as slipping whereas if the movement was smaller than 1/2 pixel, it was considered non-slipping or 194 stuck to the glass. The stick ratio, the ratio of non-slipping area to the total contact area, was obtained for each 195 pair of consecutive frames. 196

Statistical analyses 197
All statistical analyses were performed in MATLAB, using the functions corr (for Pearson correlation), ttest (for 198 paired t-tests) and regress (for linear regression). The power-law fits were obtained by computing the 199 coefficients of a linear regression on the logarithmic transformation of the data. 200

201
We asked subjects to perform paced vertical oscillations holding the manipulandum in precision grip, i.e. with 202 the thumb and index finger (Fig 2B). Figure 3 shows the evolution of the recorded variables during a typical 203 experimental block. A video of the fingerpad during a typical trial is also provided in the supplementary 204 materials. The manipulandum weight was fully compensated by a counterweight, and therefore LF was only due 205 to the object acceleration and was zero when the object was still. Thus, as expected, the manipulandum 206 movement (Fig 3A) generated cyclic fluctuations in LF due to the inertial forces, and we observed two peaks of 207 load force that were very similar in amplitude, one in the upper and one in the lower part of the trajectory. The 208 load force variations were accompanied by variations of the grip force applied by the subjects (median 209 correlation across oscillations, r=0.61±0.17, mean±std across subjects, n=18, see also herebelow and Figure 6), 210 and LF peaks gave rise to synchronized grip force peaks ( Fig 3B). These GF peaks were such that the GF/LF ratio 211 was kept above the friction limit in most cases (in red, Fig 3C), and therefore the object did not slip. However, 212 even though full slip was avoided during most oscillations, the zone of stable contact also fluctuated, as partial 213 slip was observed and was quantified by the stick ratio (SR, Fig 3D), which typically reached a minimum when 214 the constraints were maximal, that is at the maximum of LF or the minimum GF to LF ratio. As described earlier 215 in a passive context (Delhaye et al. 2016), partial slip is accompanied by significant strain patterns in the slipping 216 regions. The heatmaps provided in the bottom part of Figure 3 show the evolution of the three independent 217 components of the strain rate tensor at each point of the contact area during one oscillation (Fig 3E). The last 218 line shows the norm of the strain rate tensor (Fig 3F, see Methods). While it does not bear any physical meaning, 219 the strain rate norm provides a clear picture of the distribution of the strains and their intensities inside the 220 contact area, irrespective of the type of deformation and invariant to rotation. As expected from previous 221 reports, the strains take place at the periphery of the contact and propagate further toward the center as the SR 222 decreases, i.e. when the GF to LF ratio comes closer to the friction limit. As shown in a typical trace in Figure  Since all signals followed a stereotyped pattern for each oscillation, we looked at the averaged evolution of the 236 traces during all oscillations (Figure 4). The GF, LF, and SR patterns follow the behavior described earlier (Fig 4A-237 B, see also Figure 3). That is, the maximum GF and the minimum SR were observed on average at the time of the 238 maximum LF. We also looked at the vertical displacement of the center of pressure (COP, Fig 4C), which is 239 mainly caused by a redistribution of the pressure inside the contact area and the rolling of the finger. We 240 observed that the COP moved significantly downward during the first half of the oscillation and upward during 241 the second half of the oscillation. Even though each oscillation showed a slightly different pattern of 242 deformation and widely different levels of minimum SR across subjects (see below), the general shape follows a 243 trend that is expected from previous work (Delhaye et al 2016). The strain rate amplitude, as measured by the 244 90 th percentile of the strain rate norm across the entire contact, typically followed a trajectory similar to the 245 variation of LF, or LF rate. Fig 4E shows the averaged pattern of strain rate during one oscillation for a typical 246 subject. Since the central stuck zone is pulled up by the glass during the upper part of the trajectory (and pulled 247 down during the lower part of the trajectory), the vertical strains (e_yy) were tensile in the lower regions of the 248 contact area and compressive in the upper region of the contact area ( Fig 4E). As a consequence of the elastic 249 properties of the skin, horizontal strains (e_xx) appeared following the same pattern as the vertical ones, except 250 that they were tensile where the vertical strains were compressive and vice versa. They were also smaller in 251 amplitude. Finally, shear strains were substantial on the proximal (left) and distal (right) parts of the contact 252 area. The norm of the strain rate (e_n) shows the areas of the skin where the strains are the largest, irrespective 253 of the direction of the strains. It can be observed from those that a wave of strains progresses from the exterior 254 of the contact area towards its center as the stick ratio decreases. In Figure 3, the wave of strains doesn't reach 255 the center of the fingertip as the stick ratio doesn't reach zero. Looking at subjects' gripping behavior (Figure 5), as described by the average GF value during one oscillation, we 268 found that it was consistent across the entire experiment ( Fig 5A). Indeed, even though the average grip force 269 was on average slightly higher in the first block, this difference was not significant when compared to the last 270 block (Fig 5C, paired t-test, t(17) = -1.14, p = 0.272). Within a block, there was a clear tendency of subjects to 271 progressively decrease their grip force (Fig 5A). This was not related to a change in the movement kinematics, 272 which should affect the load force ( Fig 5B). Indeed, even though the first two oscillations of each block seemed 273 to sometimes generate a higher load force (i.e. faster movement), the average load force of the third movement 274 was already at the same level as the last movement (Fig 5D, right, paired t-test, t(17) = -1.59, p = 0.130). 275 Interrestingly, GF progressively decreased during the experimental block and the difference between the grip 276 force of the third and the last movement was highly significant (Fig 5D,  While the skin deformation patterns were similar across subjects and the grip force levels were constant across 279 the experiment, we observed that different subjects used very different levels of grip force leading to very 280 different amplitudes of deformations. Indeed, we found that the subjects that were instructed to use a minimal 281 GF tended to apply an amount of GF just above the minimum required by friction, as shown by the peak in GF 282 being just above the slip force estimated thanks to the friction measurements ( Fig 5E). However, other subjects 283 applied an excessive amount of GF. This is not resulting from the subject not coordinating their GF with the LF, 284 as the subjects applying excessive GF also showed a high level of GF modulation, thereby taking into account the 285 LF modulation (Fig 5F). It was rather explained by a high level of mean GF (Fig 5F). The different levels of GF that individual subjects used to perform the oscillations led to different observations 300 about the parameters of the manipulation ( Figure 6). First, the vertical displacement of the COP (see Methods) 301 tended to strongly decrease with the grip force level (Fig 6A). This decrease was observed across subjects, with a 302 relationship that followed a negative power law (Fig 6A left, R 2 =0.91, F(1,16)=165.65, p<0.001, ( ) = . , 303 with a=10.93 and b=-0.84). It was also observed within-subjects, with a negative correlation that was significant 304 at the group level (Fig 6A, right, paired t-test, t(17) = -4.45, p<0.001). This trend was not followed for the two 305 subjects with the lower GF values, probably because of the very small range of GF (as shown by the very small 306 horizontal wiskers for the two blue dots in Fig 6A right). The skin displacement range was also quantified. It 307 measures the maximal vertical range of motion of a fingerprint feature inside the contact area and therefore 308 quantifies how much the skin moves within the slipping regions (Fig 6B, left). This variable is strongly correlated 309 with the maximum level of skin strain rate (shown in Fig 6C, correlation r=0.88). Both variables decreased with 310 the grip force, according to a power-law (Fig 6B,  . , with a=84.37 and b=-1.04), and the negative correlation within subjects was also observed (Fig 6B, right,  313 displacement range: paired t-test, t(17) = -3.55, p = 0.002; Fig 6C, right, strain rate norm: paired t-test, t(17) = -314 3.32, p = 0.004). Again, the within-subjects correlation was lower for some subjects having small variations of GF 315 across trials. 316 The maximal strain rates were substantial, and ranged from 6%/s to 100%/s (Fig 6D, left). Obviously, the 317 maximal strain rates strongly correlated with the maximal strain values observed during an oscillation (r=0.99), 318 which ranged from 1% to 17% depending on the subject's GF level. Finally, as could be expected, the minimum 319 value of the stick ratio within an oscillation was low for low levels of grip force and tended to 1 for very high 320 levels of grip force (Fig 6D, left). We did not observe any significant tendency at the group level in the within-321 subjects correlation (Fig 6D, right, paired t-test, t(17) = 1.60, p = 0.128). As a control, we verified that the use of 322 a high level of grip force was not related to the level of load force, but rather to an inappropriate adjustment to 323 friction. We found that indeed, while there were some variations in the level of load force across subjects, 324 related to variations in the movement acceleration from trial to trial, the correlation with the grip force was low 325 (r=0.17, Fig 6E, left). However, within-subjects, we observed a clear positive correlation at the group level, 326 meaning that the subjects adapted their level of grip force to the movement kinematics ( Fig 6E, right, paired t-327 test, t(17) = 5.74, p<0.001). As could be expected, this tendency was strongest for the low GF values, for which 328 the risk of slip is higher, and therefore a tight coupling between the LF and GF is required. In summary, we showed that higher levels of GF yielded less fingertip rolling (as measured with the vertical 341 displacement of the center of pressure), fewer skin strains (as measured with the strain rate norm) and a higher 342 proportion of the skin remaining stuck on the glass (as measured with the stick ratio). 343

344
In this paper, we described a novel device that enables synchronous monitoring of strains in the fingertip skin in 345 contact with a manipulated object and the forces exerted by the fingers on the object. The device was tested in 346 an experiment involving 18 subjects and permitted to describe and quantify the strain patterns emerging from 347 active object manipulation. This is the first study showing that substantial skin strain rates (>50%/s) take place 348 inside the contact area during active manipulation, even though the object is in a stable, non-fully slipping 349 contact and the object never dropped from the hand of the subjects. 350 The development of this device was inspired by a previous publication presenting a much simpler apparatus 351 (Tada et al. 2002). This equipment was also manipulated in a precision grip and was also comprised of force 352 sensors under each finger and of a camera to monitor the stick ratio, which relied on 167 manually drawn points 353 on the fingertip to monitor skin displacement from frame to frame. Importantly, our setup has a much higher 354 resolution (64 pixels per mm) and therefore enables an accurate measurement of local skin strains, which was 355 not possible earlier. 356 Given the substantial amplitude of the strain rates measured in this study, it is reasonable to assume that those 357 strains can be faithfully captured by the tactile afferents, and in particular the FA-I afferents (fast adapting type I 358 afferents). Indeed, we recently demonstrated, in a passive setup, that FA-I afferents are sensitive to the 359 compressive strain rates related to partial slip, comparable in scale to those observed here (Delhaye et al. 2021). 360 Those strains therefore likely provide tactile feedback about grip safety. 361 There was a wide range of mean grip force levels used by different subjects observed in this study. Some 362 subjects tended to apply a grip force level just above the friction limit while some others exerted a much higher 363 grip force. Those behaviors lead to very different feedback. The "safe" strategy -exerting a high level of grip 364 force -leads to a very limited amount of partial slip and a low level of deformation. Therefore, it also provides a 365 limited amount of information related to contact stability. The "risky" strategy -using a grip force level just 366 above the friction limit -leads to a large fraction of the contact area in partial slip and substantial strain rates 367 inside the contact. Therefore, this strategy enables rich tactile feedback about the contact state at each 368 oscillation. 369 While there was a clear relationship between the level of grip force and the amount of skin strains (Fig 6C), 370 there was also a large variability. This variability can be explained by individual differences in skin properties. 371 Indeed, the fingerpad can have very different mechanical and geometrical properties: stiffness (Wang and  372 Hayward 2007), humidity (André et al. 2010), size (Peters et al. 2009), and all of those likely influence 373 deformation and perception (Gueorguiev et al. 2016;Peters et al. 2009). 374

Limitations 375
The developed device has some limitations. First, for a device that is intended to be used in precision grip, that is 376 pinched between the thumb and index finger, its mass is rather large (around 500 grams). This limitation can be 377 compensated by a counterweight, as done in this work, but it modified the weight/inertia relationship of the 378 object, which makes it more remote from natural object manipulation. The manipulation of an object with no 379 weight but high inertia, as done in the present study, is comparable to object manipulation in weightlessness in 380 terms of contact forces with the object. Furthermore, due to the counterweight setup, we are limited to vertical 381 movements. In addition, we used smooth transparent glass. While this is a very convenient material for imaging 382 the skin-object contact, the flat contact with a glass surface is very different from the rough contact experienced 383 with most natural objects. 384 As summarized above, some subjects tended to use a high level of grip force, much higher than the friction limit. 385 This constrats with previous work showing that, under normal sensory feedback, people tend to manipulate 386 objects with a grip force close to the slip limit (Augurelle et al. 2003). This might be explained by three factors. 387 First, the instrumented object might look fragile and subjects tended to apply excessive GF levels to make sure 388 not to drop it. Second, by compensating the weight of the object with a counterweight (see Methods), the 389 experienced weight is much lower than the one expected from the object's appearance and size, which may 390 surprise the subjects. Moreover, there is a discrepancy between the perceived weight of the object, which is 391 zero, and the interaction force resulting from the doubled inertia. As a result, it is likely that an excessive GF is 392 exerted by caution. Finally, glass remains an material having particular frictional properties, with a coefficient of 393 friction that can vary by up to one order of magnitude depending on several factors (Adams et al. 2013;Delhaye 394 et al. 2014;Pasumarty et al. 2011), and might therefore encourage higher safety margin even if not needed. 395

396
The device developed in this study will enable the monitoring of fingertip skin strains at the finger-object contact 397 and provide a window into the feedback from tactile afferents inside the contact area during manipulation.

398
Future work will test if unexpected changes in the object parameters, such as a change in friction, can be quickly 399 detected and accounted for thanks to the feedback provided by partial slip. 400