betaFeedback

WorksComputer Science

Vortex-Induced Torque in Grasping Soft Objects: Influence of Geometry and Pneumatic Parameters

Abstract

Vortex grippers are commonly integrated into automated systems for handling objects with complex geometries in packaging, sorting, and assembling. However, their interaction with compliant objects remains difficult to predict, particularly with respect to the vortex-induced torque transmitted during grasping. This study investigates vortex-induced torque during the grasping of compliant objects and quantifies the coupled effects of supply pressure, tangential nozzle diameter, surface geometry, and curvature. The results show that increasing the tangential nozzle diameter primarily enhances the sensitivity of the generated torque to supply pressure rather than producing a uniformly higher torque. Surface geometry also strongly influences the rotational interaction, with dome-concave surfaces generally producing substantially greater torque than corresponding dome-convex surfaces, whereas the difference between cylindrical concave and convex surfaces is comparatively small. The strongest and most repeatable response was observed for a dome-concave surface exhibiting favorable object-gripper geometric matching. Among the evaluated machine-learning regression methods, Gaussian Process Regression provided the best overall predictive performance under repeated cross-validation, demonstrating that vortex-induced torque can be estimated from combined pneumatic and geometric parameters. These findings provide a basis for predicting and utilizing vortex-induced torque in the dexterous manipulation of soft objects.

Introduction

Pneumatic gripping devices are very easy to integrate into robotic automation processes due to their predictability and ease of use [1, 2]. This is a general statement that is only partially true. There are many different subtypes of pneumatic gripping devices [3, 4]: vacuum [5, 6, 7] (pumpless, piston, ejector, pumping); air-jet [8, 9, 11] [10] (vortex, Bernoulli, combined); combined [12, 13, 14, 15] (pneumatic-pneumatic, pneumatic-other). The complexity of their integration into automated systems is that most of them are built on different physical principles. This does not easily allow for predicting the behavior of the gripper with different objects, especially soft ones.

By considering the most widely used suction cups based on ejector systems, it is obvious that the behavior of the object can be easily predicted by uniform distribution of air pressure by the suction cup over the surface of the object [16, 17]. However, when it comes to air-jet grippers, the interaction between the gripper and the object becomes very complex [10]. This can be seen if we do not consider soft objects but only rigid objects; changing the parameters of the gripper geometry [9], nozzle [19], object geometry [20], supply pressure [21, 22], and object-gripper gap [23, 24, 25] leads to significant changes in both the object deformation and the force characteristics of the gripper. This is due to the fact that the pressure distribution on the surface of the object will not be uniform (Figure 1a), and zones of negative and positive pressure will be formed [26, 27].

Vortex gripper technology and representative applications

Figure 1. Vortex gripper technology and representative applications: (a) compressed air flow direction that form a vortex rotation flow that creates a lifting effect [3]; (b) application of vortex technology for grasping objects in manufacturing and assembly (electronics, food, packaging, etc.) [33]; (c) integration of soft vortex gripper into the dexterous hand finger [44]; (d) vortex gripper for grasping soft tissue of various shapes [49, 50].

Such characteristics of jet grippers have both advantages and disadvantages; for example, they allow grasping an object from a distance [28, 29, 30], but they create additional vibrations [31]. Due to the ability of vortex and Bernoulli grippers to hold objects that are not completely connected to the gripper (like suction cups failing under these conditions), they have gained popularity for grasping fragile [32], porous [11], or electronic parts [33, 34] (Figure 1b). Additionally, the ability to grasp objects from a distance [35] has expanded the scope of applications of jet grippers for deformable [36, 37, 38] and soft [39] objects. They are most widely used for grasping textile objects from a stack [11, 40] and for local control of the position of deformable materials in manufacturing [41, 42]. Many studies have focused on the use of jet grippers for both dexterous manipulation through their integration into the fingers [43, 44] of anthropomorphic hands (Figure 1c) and for dexterously manipulating deformable objects [45] with the robot arm.

However, unlike Bernoulli grippers, vortex grippers have additional characteristics that positively affect the lifting force due to a greater ability to adapt to the surface of the object [46], as well as negatively, which leads to the rotation of the object during grasping. This is because the lifting force is created by the rotation of the airflow, forming a vortex that lifts the object but also creates a torque that rotates the object around the axis of the gripper (Figure 1a). Researchers and developers of vortex grippers usually avoid discussing this negative factor and solve the problem by integrating friction elements, which also works for other jet grippers [47]. Also, if it is not possible to use friction elements, side stops are integrated to prevent the object from rotating. By using wheels as friction elements, we can see the application of vortex technology to mobile robots moving along walls [48].

Despite the torque effect, vortex grippers have many positive effects that allow them to be used for soft tissue grasping in medical applications. In our previous studies [49, 50], we observed the rotation of bio-inspired tissue during vortex grasping and mitigated this effect using friction elements (Figure 1d). However, the influence of compliant-object geometry on vortex-induced torque remains insufficiently characterized, limiting the ability to predict rotational interaction during soft-object grasping. A quantitative understanding of this torque may also enable future manipulation strategies in which the vortex-induced rotational effect is deliberately exploited rather than suppressed. It is also important to understand that knowledge of torque parameters can enable in-hand manipulation without dexterous movements, using only vortex flows. Therefore, in this paper, a detailed study of vortex-induced torque in grasping soft objects was conducted, with 4920 valid trials distributed over 492 experimental conditions. The main contributions of this study are summarized as follows:

  • Experimental quantification of vortex-induced torque as a function of supply pressure, tangential nozzle diameter, surface geometry, and curvature was conducted.

  • Identification of geometry-dependent torque amplification for dome-concave surfaces, including a highly repeatable local maximum near a specific object-gripper geometric relationship.

  • Demonstration that lifting force and vortex-induced torque exhibit different dependencies on the gripper-surface gap, showing that rotational interaction cannot be inferred from lifting force behavior alone.

  • Introduction of an object-gripper geometric-matching interpretation to describe the observed torque amplification and the differences between dome and cylindrical surfaces.

  • Development and cross-validated comparison of machine-learning regression models for predicting peak vortex-induced torque from pneumatic and geometric parameters, with Gaussian Process Regression providing the best overall performance.

Methodology

Experimental Setup

The experiments were conducted to quantify the rotational interaction generated by the vortex flow during gripper contact with compliant surfaces and to determine how this interaction is influenced by supply pressure, tangential nozzle diameter, and local surface geometry. The primary response variable was the torque about the vortex axis, MzM_z, whereas the axial force FzF_z was additionally used to identify the transition from mechanically compressed contact to the vortex lifting stage (Figure 2).

Experimental setup with different object surfaces for studying the force characteristics of a vortex gripper

Figure 2. Experimental setup with different object surfaces for studying the force characteristics of a vortex gripper.

The experimental setup (Figure 2) consists of a 7-DOF robot arm with an attached gripper control unit and an end-effector adapter. A vortex gripper is attached to the end-effector adapter via an ATI Nano17 force/torque sensor with an accuracy of ±0.25%\pm0.25\%. Three geometrically identical vortex grippers were investigated. The grippers had an outer diameter of Dg=20D_g = 20 mm, a vortex-chamber diameter of Dc=14D_c = 14 mm, and a chamber depth of Hc=4H_c = 4 mm. Each gripper contained two tangential nozzles. The investigated nozzle diameters were dnd_n in 0.6, 0.8, 1.0 mm for corresponding gripper configurations G1G_1, G2G_2, and G3G_3, respectively. Four supply pressure values, PsP_s in 100, 200, 300, 400 kPa were investigated. The compliant test represented four curved surface types: cylindrical concave, cylindrical convex, dome concave, and dome convex. Ten nominal surface radii were studied,

The experimental setup (Figure 2) consists of a 7-DOF robot arm with an attached gripper control unit and an end-effector adapter. A vortex gripper is attached to the end-effector adapter via an ATI Nano17 force/torque sensor with an accuracy of ±0.25%\pm0.25\%. Three geometrically identical vortex grippers were investigated. The grippers had an outer diameter of Dg=20D_g = 20 mm, a vortex-chamber diameter of Dc=14D_c = 14 mm, and a chamber depth of Hc=4H_c = 4 mm. Each gripper contained two tangential nozzles. The investigated nozzle diameters were dnd_n in 0.6, 0.8, 1.0 mm for corresponding gripper configurations G1G_1, G2G_2, and G3G_3, respectively. Four supply pressure values, PsP_s in 100, 200, 300, 400 kPa were investigated. The compliant test represented four curved surface types: cylindrical concave, cylindrical convex, dome concave, and dome convex. Ten nominal surface radii were studied, RR in 15, 20, 25, 30, 35, 40, 45, 50, 75, 100 mm. A compliant flat surface was additionally investigated. Thus, each gripper configuration was evaluated on 41 different surface geometries. The interaction layer was based on standardized bio-inspired silicone soft tissue, which was fabricated from two-component silicones (Smooth-On) and the two different layers, representing skin and subcutaneous fat [50].

During each surface experiment, the gripper was first brought into contact with the soft surface and pressed against it. Compressed air was subsequently supplied while the gripper-surface interface remained initially closed. The gripper was then moved vertically upward at a constant velocity of vz=10v_z = 10 mm/s, thereby opening the gap between the gripper and the compliant surface. Ten repetitions were performed for every combination of nozzle diameter, pressure, and surface geometry. The complete nominal factorial design therefore contained 4920 individual trials.

The measured MzM_z (Figure 2) represents the reaction torque acting on the gripper rather than a direct measurement of the torque transmitted exclusively to the grasped object. The measured signal can additionally contain contributions from the free airflow, structural loading, and other quasi-static effects. For this reason, the largest raw ∣Mz∣\lvert M_z\rvert value recorded over the complete robot trajectory was not used directly. The force and torque signals were acquired at a sampling frequency of fs=2000f_s = 2000 Hz. The sign convention was selected such that negative FzF_z represents mechanical compression of the gripper, whereas positive FzF_z corresponds to the direction of the lifting force. The beginning of the analyzed vortex-induced torque was determined directly from the raw FzF_z signal, where i0i_0 is the first sample satisfying Fz(i0−1)<0F_z(i_0 - 1) < 0, Fz(i0)≥0F_z(i_0) \ge0.

Statistical Analysis

For each combination of gripper configuration, pressure, and surface geometry, the mean peak torque was calculated from the nn available repetitions as

M‾z,peak=1n∑j=1nMz,peak(j).(1)\overline{M}_{z,\mathrm{peak}} = \frac{1}{n}\sum_{j=1}^{n} M_{z,\mathrm{peak}}^{(j)}. \tag*{(1)}

Trial-to-trial variability was represented by the sample standard deviation,

σMz=1n−1∑j=1n(Mz,peak(j)−M‾z,peak)2.(2)\sigma_{M_z} = \sqrt{\frac{1}{n-1}\sum_{j=1}^{n}\left(M_{z,\mathrm{peak}}^{(j)}-\overline{M}_{z,\mathrm{peak}}\right)^2}. \tag*{(2)}

The coefficient of variation was additionally calculated as CV=100(σMz/M‾z,peak)CV = 100(\sigma_{M_z}/\overline{M}_{z,\mathrm{peak}}). To summarize the effect of pressure while reducing distortion from very large percentage changes, a logarithmic pressure-sensitivity metric was defined as

SP=log⁡2(M‾z,peak(400 kPa)M‾z,peak(100 kPa)).(3)S_P = \log_2\left(\frac{\overline{M}_{z,\mathrm{peak}}(400\ \mathrm{kPa})}{\overline{M}_{z,\mathrm{peak}}(100\ \mathrm{kPa})}\right). \tag*{(3)}

This metric has a direct interpretation:

RR [mm]152025303540455075100
δ(7,R)\delta(7,R) [mm]1.7341.2651.0000.8280.7070.6170.5480.4920.3270.245
δ(10,R)\delta(10,R) [mm]3.8202.6792.0871.7161.4591.2701.1251.0100.6700.501
Λg\Lambda_g0.9550.6700.5220.4290.3650.3180.2810.2530.1670.125

Table 1. Geometric Deviation of the Ideal Dome Relative to the Gripper Dimensions

SP={0,no change,+1,twofold increase,−1,twofold decrease.(4)S_P = \begin{cases} 0, & \text{no change},\\ +1, & \text{twofold increase},\\ -1, & \text{twofold decrease}. \end{cases} \tag*{(4)}

Similarly, the influence of curvature orientation was quantified by matching the concave and convex surfaces at identical surface shape, radius, pressure, and nozzle diameter:

Sc/v=log⁡2(M‾z,peakconcaveM‾z,peakconvex).(5)S_{c/v} = \log_2\left(\frac{\overline{M}_{z,\mathrm{peak}}^{\mathrm{concave}}}{\overline{M}_{z,\mathrm{peak}}^{\mathrm{convex}}}\right). \tag*{(5)}

Thus, Sc/v>0S_{c/v} > 0 indicates greater torque for the concave geometry, while Sc/v<0S_{c/v} < 0 indicates greater torque for the corresponding convex geometry. Where monotonic radius trends were investigated using the Spearman correlation coefficient:

ρs=corr⁡[rank⁡(R),rank⁡(M‾z,peak)].(6)\rho_s = \operatorname{corr}\left[\operatorname{rank}(R),\operatorname{rank}\left(\overline{M}_{z,\mathrm{peak}}\right)\right]. \tag*{(6)}

The final analysis contained 4920 valid trials distributed over 492 experimental conditions. Despite the compliant nature of the specimens and the transient character of the interaction, the measurements exhibited generally good repeatability. The coefficient of variation was below 10% for 26.0% of the investigated conditions, below 20% for 69.3%, below 30% for 88.0%, and below 50% for 97.4%.

Object-Gripper Relationship

For a spherical surface of radius RR, the geometric deviation from the tangent plane at radial position rr can be expressed as δ(r,R)=R−R2−r2\delta(r,R) = R - \sqrt{R^2-r^2}. Two radial positions are particularly relevant to the present gripper: rc=Dc/2=7r_c = D_c/2 = 7 mm, rg=Dg/2=10r_g = D_g/2 = 10 mm, corresponding to the vortex-chamber radius and the outer gripper radius, respectively.

To express the magnitude of the surface curvature relative to the chamber depth, the dimensionless geometric ratios were defined as Λc(R)=δ(rc,R)/Hc\Lambda_c(R) = \delta(r_c,R)/H_c, Λg(R)=δ(rg,R)/Hc\Lambda_g(R) = \delta(r_g,R)/H_c (Table I). For a comparatively small peripheral gap, the available annular outlet area scales approximately as Aout∼2πrgg(rg)A_\mathrm{out} \sim2\pi r_g g(r_g), where g(rg)g(r_g) denotes the local peripheral gap at the outer gripper radius. Consequently, relatively small changes in the peripheral gap can produce substantial changes in the effective airflow outlet area.

At R=15R = 15 mm, the geometric deviation at the outer gripper radius is δ(10,15)=3.820\delta(10,15) = 3.820 mm, which is close to the complete vortex-chamber depth of 4 mm. The corresponding dimensionless value is Λg(15)=0.955\Lambda_g(15) = 0.955. At R=20R = 20 mm, δ(10,20)=2.679\delta(10,20) = 2.679 mm, and Λg(20)=0.670\Lambda_g(20) = 0.670. Thus, the R=15R = 15-20 mm surfaces modify the gripper-surface geometry on the same dimensional scale as the vortex chamber itself.

ML for Torque Prediction

Peak vortex-induced torque was predicted using 492 condition-level means, each obtained from ten repetitions; individual trials were not treated as separate ML samples. The input vector was

x=[p, An, Λc±, Nκ]T,(7)\mathbf{x} = [p,\ A_n,\ \Lambda_c^{\pm},\ N_\kappa]^{\mathsf{T}}, \tag*{(7)}

where p=Ps/(400 kPa)p = P_s/(400\ \mathrm{kPa}), An=πdn2/2A_n = \pi d_n^2/2 is the total nozzle area in mm2\mathrm{mm}^2, and Nκ=0,1,2N_\kappa= 0,1,2 for flat, cylindrical, and dome surfaces, respectively. For curved surfaces,

Λc±=sR−R2−rc2Hc,s={+1,concave,−1,convex,(8)\Lambda_c^{\pm} = s\frac{R-\sqrt{R^2-r_c^2}}{H_c},\qquad s = \begin{cases} +1, & \text{concave},\\ -1, & \text{convex}, \end{cases} \tag*{(8)}

with rc=7 mmr_c = 7\ \mathrm{mm} and Hc=4 mmH_c = 4\ \mathrm{mm}; Λc±=0\Lambda_c^{\pm}=0 for flat surfaces.

Five models were compared: GPR, Random Forest (RF), Extra Trees, AdaBoost, and an equally weighted RF–AdaBoost ensemble. RF and Extra Trees used 700 trees, unrestricted depth, minimum split/leaf sizes of 2/1, and all features at each split; bootstrap sampling was enabled only for RF. AdaBoost used up to 500 trees of maximum depth 4, minimum leaf size 2, learning rate 0.03, and square loss. These settings were fixed without a hyperparameter search.

GPR used a constant-scaled Matérn-5/2 kernel with four feature-specific length scales and additive white noise. Initial values (bounds) were 1 (10−310^{-3}–10310^3) for the constant, 1 (10−210^{-2}–10210^2) for each length scale, and 0.02 (10−610^{-6}–10110^1) for noise variance, with diagonal regularization of 10−810^{-8}. Kernel parameters were optimized by maximizing the training-fold log marginal likelihood without optimizer restarts. GPR inputs and targets were standardized using training-fold statistics, and predictions were returned to torque units. Tree-based models used no additional standardization.

Evaluation used repeated five-fold cross-validation with five repetitions, identical splits for all models, and random seed 42 for splitting and stochastic estimators. Nested cross-validation was not used. All fitting and GPR preprocessing were performed within each training fold. The R2R^2, RMSE, and MAE were reported as mean ±\pm standard deviation across 25 validation folds; this standard deviation is not a confidence interval. The model with the highest mean R2R^2 was selected.

Results and Discussion

Rigid Surface Quasi-Static Interaction

A rigid-surface reference experiment was conducted to independently characterize the dependence of lifting force and vortex-induced torque on the gripper-surface gap when surface deformation was eliminated (Figure 3). For this, the G1_1 gripper with 0.6 mm tangential nozzles and a rigid flat surface was used. The gripper was positioned at discrete stationary gap heights h=0, 0.05, 0.10,…,1.00 mmh = 0,\ 0.05,\ 0.10,\ldots,1.00\ \mathrm{mm}. At every height, 100 force/torque sensor readings were taken, and their mean was recorded as the response at the corresponding gap height. The procedure was repeated for supply pressures of 100, 200, 300, and 400 kPa. The transition from mechanically compressed contact to positive lifting force was characterized by the gap height h0h_0 at which Fz(h0)=0F_z(h_0)=0. For adjacent measured heights hkh_k and hk+1h_{k+1} satisfying Fz(hk)<0F_z(h_k)<0, Fz(hk+1)≥0F_z(h_{k+1})\geq0, the zero-crossing height was estimated using linear interpolation:

Quasi-static lifting force and torque magnitude as functions of gap height for the rigid flat surface using G1

Figure 3. Quasi-static lifting force FzF_z and torque magnitude ∣Mz∣|M_z| as functions of gap height for the rigid flat surface using G1_1.

h0=hk−Fz(hk)hk+1−hkFz(hk+1)−Fz(hk).(9)h_0 = h_k - F_z(h_k)\frac{h_{k+1}-h_k}{F_z(h_{k+1})-F_z(h_k)}. \tag*{(9)}

The maximum lifting force was defined as Fz,max⁡rigid=max⁡hFz(h)F_{z,\max}^{\mathrm{rigid}}=\max_h F_z(h), and the magnitude of the rotational response was calculated as Mz,rigid(h)=∣Mz(h)∣M_{z,\mathrm{rigid}}(h)=|M_z(h)|. The corresponding mean torque was calculated as

M‾z,rigid=1Nh∑q=1Nh∣Mz(hq)∣.(10)\overline{M}_{z,\mathrm{rigid}}=\frac{1}{N_h}\sum_{q=1}^{N_h}|M_z(h_q)|. \tag*{(10)}

The transition from mechanical compression to positive lifting force occurs at progressively greater gaps as the supply pressure increases. The interpolated zero-crossing heights are 0.070, 0.079, 0.085, 0.110 mm for 100, 200, 300, and 400 kPa, respectively. The maximum occurs at h=0.20h=0.20 mm for 100 kPa and at h=0.25h=0.25 mm for 200, 300, and 400 kPa.

In Table II, ∣Mz∣|M_z| represents the mean torque magnitude over the developed-gap interval 0.30≤h≤1.000.30\leq h\leq1.00 mm. The maximum lifting force increases from 0.696 N at 100 kPa to 2.685 N at 400 kPa. Over the four investigated pressure levels, a linear regression gives Fz,max⁡rigid≈0.00665Ps−0.0248F_{z,\max}^{\mathrm{rigid}}\approx0.00665P_s-0.0248, with R2=0.995R^2=0.995. This experiment is used only as a compact description of the measured pressure range and should not be interpreted as a general physical model of the vortex gripper. The torque shows a substantially different dependence on gap height. Once the gap exceeds approximately 0.30 mm, the torque magnitude becomes nearly independent of further increases in height. The fitted slopes over 0.30≤h≤1.000.30\leq h\leq1.00 mm are only 0.028, -0.011, 0.006, 0.005 N·mm/mm at 100, 200, 300, and 400 kPa, respectively.

PsP_s [kPa]h0h_0 [mm]Fz,max⁡F_{z,\max} [N]h(Fz,max⁡)h(F_{z,\max}) [mm]∣Mz∣|M_z| [N·mm]
1000.0700.6960.200.319 ± 0.008
2000.0791.2430.250.560 ± 0.008
3000.0851.9310.251.304 ± 0.005
4000.1102.6850.251.796 ± 0.005
dnd_n100 kPa200 kPa300 kPa400 kPa
0.6 mm0.238 ± 0.0480.212 ± 0.0220.307 ± 0.1100.282 ± 0.064
0.8 mm0.620 ± 0.0500.757 ± 0.0800.952 ± 0.0610.963 ± 0.110
1.0 mm0.541 ± 0.0960.715 ± 0.2180.994 ± 0.2671.601 ± 0.376

Table II. Quasi-Static Rigid-Flat Surface Interaction

Consequently, Fz(h)F_z(h) and Mz(h)M_z(h) show fundamentally different gap dependencies. The lifting force has a distinct optimum gap near 0.20–0.25 mm and subsequently decreases with increasing gap. In contrast, once the vortex provides a lifting force, the rotational response remains approximately constant over a considerably wider gap interval.

Impact on Torque Sensitivity

The flat compliant surface provides the simplest geometry for evaluating the coupled effects of nozzle diameter and supply pressure. Figure 4a presents the mean peak torque as a function of supply pressure. The numerical results are summarized in Table III.

Figure 4. (a) Vortex-induced torque for the flat compliant surface as a function of supply pressure. (b) Logarithmic pressure sensitivity of the torque between 100 and 400 kPa (SP=0SP = 0 indicates no change, SP=1SP = 1 represents a twofold increase, and SP=−1SP = -1 represents a twofold decrease).

With the 0.6 mm nozzles, the vortex-induced torque remained comparatively small and showed no clear monotonic dependence on pressure. The mean torque was 0.238 ± 0.048 N·mm at 100 kPa and 0.282 ± 0.064 N·mm at 400 kPa, corresponding to an increase of only 18.6%. A substantially stronger pressure dependence was obtained when the nozzle diameter was increased. For the G2 (dn=0.8d_n=0.8 mm) configuration, the mean torque increased from 0.620 ± 0.050 N·mm at 100 kPa to 0.963 ± 0.110 N·mm at 400 kPa, representing an increase of 55.3%. The strongest pressure dependence was obtained using the G3 gripper. The corresponding torque increased from 0.541 ± 0.096 N·mm to 1.601 ± 0.376 N·mm, corresponding to an increase of approximately 196%. The results indicate that increasing nozzle diameter does not simply increase torque by a constant factor. Instead, one of its most important effects is to increase the sensitivity of the generated torque to the supplied pressure.

The torque-dependent behavior observed for the flat surface was also reflected across the complete geometry set. Considering all 41 investigated surfaces, the median changes in mean peak vortex-induced torque between 100 and 400 kPa were 23.6%, 53.7%, 111.6% for the G1, G2, and G3 grippers, respectively. Torque increased from 100 to 400 kPa in 25 of 41 geometries for G1, 34 of 41 geometries for G2, and 38 of 41 geometries for G3. The torque dependence

Geometry0.6 mm ΔMz\Delta M_z0.6 mm Positive0.8 mm ΔMz\Delta M_z0.8 mm Positive1.0 mm ΔMz\Delta M_z1.0 mm Positive
Cylindrical concave−9.5%3/1035.8%7/10113.7%10/10
Cylindrical convex50.8%8/1039.4%7/10109.4%10/10
Dome concave19.5%6/1057.6%9/1063.2%7/10
Dome convex91.5%7/1065.7%10/10140.0%10/10

Table 4. Median Change in Mean Peak Torque Between 100 and 400 kPa

is summarized using the logarithmic metric of Eq. (3) in Figure 4b. The logarithmic representation was selected because it preserves all experimental conditions while reducing visual domination by very large percentage changes.

Table IV summarizes the median torque response of the four curved surface families. The nozzle-dependent effect is particularly clear for the cylindrical concave surfaces. Their median response changes from a slight reduction of 9.5% with the 0.6 mm nozzles to an increase of 113.7% with the 1.0 mm nozzles.

The results demonstrate that supply pressure, nozzle diameter, and surface geometry are strongly coupled. A larger nozzle increases the available tangential inlet area and can increase the airflow introduced into the chamber as pressure rises. However, the resulting torque additionally depends on how efficiently the rotating airflow remains coupled to the surface outlet through the gap.

Surface Influence

The radius-dependent torque for the cylindrical concave and convex surfaces can be seen in Figure 5. The cylindrical surfaces do not show a universal monotonic dependence of torque on radius. Furthermore, the difference between corresponding concave and convex cylindrical surfaces is comparatively small when all matched experimental conditions are considered.

Mean peak torque as a function of radius for cylindrical surfaces

Figure 5. Mean peak torque as a function of radius for cylindrical surfaces at different supply pressures: (a, b, c) - concave; (d, e, f) - convex.

The median concave-to-convex torque ratio is 1.06, which is close to unity and indicates that the reversal of cylindrical curvature alone does not consistently produce a large change in torque. The reason is that a cylindrical surface is curved along only one principal direction. The perpendicular direction remains locally flat, leaving portions of the peripheral airflow path comparatively unaffected by the cylinder radius. As a result, cylindrical curvature modifies the flow distribution locally without producing strong circumferential confinement around the complete gripper.

The dome geometries show a substantially stronger radius dependence (Figure 6). For dome-convex surfaces, smaller radii generally produce larger torque, particularly for the G3G_{3} gripper configuration. With the 1.0 mm nozzles, the Spearman correlation between radius and torque is negative at every pressure ρs=−0.830,−0.842,−0.842,−0.648\rho_{s} = -0.830, -0.842, -0.842, -0.648 for 100, 200, 300, and 400 kPa, respectively. The corresponding relationships are statistically significant at all four pressure levels (p<0.05p < 0.05). For example, using G3G_{3} at 400 kPa, the R=15R = 15 mm dome-convex surface produces M‾z,peak=3.285\overline{M}_{z,\mathrm{peak}} = 3.285 N·mm, whereas the R=100R = 100 mm surface produces only 0.826 N·mm.

Figure 6. Mean peak torque as a function of radius for dome surfaces at different supply pressures: (a, b, c) - concave; (d, e, f) - convex.

The dome-concave response is even more distinctive because it contains a local maximum at R=20R = 20 mm. Consequently, the measured torque cannot be represented using a simple rule such as increasing torque with decreasing radius. Instead, the data indicate a geometry-specific interaction between surface curvature and the characteristic dimensions of the vortex gripper.

The logarithmic ratio (5) compares the corresponding concave and convex surfaces present in Figure 7a. For cylindrical surfaces, the median raw ratio is only 1.06; for dome geometries, it is 1.99. Thus, the dome-concave surfaces generate approximately twice the torque of the corresponding dome-convex surfaces in the median matched condition. This effect is not produced solely by the exceptionally large R=20R = 20 mm results. When all R=20R = 20 mm pairs are excluded, the median dome ratio remains 1.74. The observed asymmetry, therefore, indicates a general influence of the three-dimensional orientation. For a cylindrical surface, one principal surface direction remains uncurved. In contrast, for a dome, the curvature modifies the gripper-surface gap around the complete circumference. A concave dome may, therefore, restrict the peripheral leakage path in all radial directions, whereas a convex dome tends to increase the peripheral gap.

Figure 7. (a) Logarithmic concave-to-convex torque ratio for matched surface radius, pressure, and nozzle diameter. (b) Mean peak torque for the R=20R = 20 mm dome-concave surface as a function of supply pressure.

The most clearly geometry-dependent response in the complete dataset occurs for the dome-concave surface with R=20R = 20 mm (Figure 7b). At every investigated pressure and nozzle diameter, the R=20R = 20 mm dome-concave geometry produces substantially greater torque than most of the remaining dome-concave surfaces. Relative to the next-highest radius under the same operating condition, the amplification ranges from approximately 1.44 to 3.92. Relative to the median of the remaining dome-concave radii, the amplification ranges from approximately 4.97 to 20.09. The maximum mean torque measured in the complete experiment is M‾z,peak=21.770±0.717\overline{M}_{z,\mathrm{peak}} = 21.770 \pm0.717 N·mm, obtained using G3G_{3} at 400 kPa. The corresponding coefficient of variation is only CV=3.29%CV = 3.29\%. The large response is therefore highly repeatable and cannot reasonably be attributed to a small number of abnormal trials.

Based on our previous vortex-gripper studies and CFD analysis [44, 50], we interpret the pronounced response at R=20R = 20 mm in terms of gap formation between the compliant object and the outer gripper surface, whose diameter is Dg=20D_{g} = 20 mm, rather than the inner vortex-chamber surface. This peripheral clearance governs the airflow discharge and may modify vortex confinement and tangential momentum transfer, providing a physical explanation for the strong geometry dependence of the measured torque. However, the specific torque maximum at R=20R = 20 mm has not been directly verified by flow-field measurements or CFD for the present configuration.

ML Prediction of Vortex-Induced Torque

Among the five evaluated regression approaches, GPR achieved the highest mean fold-level performance under repeated 5-fold cross-validation R2=0.899±0.080R^{2} = 0.899 \pm0.080, RMSE=0.508±0.217RMSE = 0.508 \pm0.217 N·mm, with an MAE=0.240±0.043MAE = 0.240 \pm0.043 N·mm. Extra Trees and Random Forest followed, with mean R2R^{2} values of 0.885±0.1010.885 \pm0.101 and 0.870±0.1300.870 \pm0.130. The Random Forest + AdaBoost ensemble and AdaBoost produced lower mean R2R^{2} values of 0.803 and 0.668. GPR was therefore selected as the final predictive model.

For the condition-level comparison shown in Figure 8, the five repeated out-of-fold predictions obtained for each condition were averaged. The resulting aggregated predictions yielded R2=0.939R^{2} = 0.939, RMSE=0.492RMSE = 0.492 N·mm, MAE=0.216MAE = 0.216 N·mm. The concentration of the predictions around the ideal 1:1 line demonstrates that the combined pneumatic and geometry-based descriptors provide sufficient information to predict the peak torque within the experimentally sampled design domain.

Figure 8. Measured and repetition-averaged out-of-fold predictions of peak vortex-induced torque for the selected GPR model.

The combined influence of surface radius and supply pressure was subsequently represented using the predicted torque map in Figure 9. The map was generated over the experimentally supported pressure interval of 100–400 kPa and the radius interval of 0–100 mm.

Figure 9. GPR prediction torque map for the vortex gripper (G3) with dome-concave surface.

The torque map clearly identifies a localized high-torque region near R≈19R \approx19–20 mm, with its magnitude increasing strongly with supply pressure. Away from this region, the predicted torque decreases substantially and varies more gradually with radius. The resulting map, therefore, provides a compact representation of the strongly coupled effects of surface geometry and pneumatic operating conditions.

A physical boundary condition Mz,peak=0M_{z,\mathrm{peak}} = 0 at Ps=0P_s = 0 was imposed in the prediction procedure. Predictions in the unmeasured pressure range 0<Ps<1000 < P_s < 100 kPa were generated using the learned relationship between pneumatic and geometric parameters, together with this zero-pressure constraint.

Overall, the machine-learning analysis demonstrates that vortex-induced torque can be represented by a continuous geometry/pressure-dependent model within the investigated design domain. The model predictions reproduce the geometry-dependent trends in the experimental dataset but do not independently validate the proposed flow mechanism. The machine-learning analysis, therefore, indicates that surface geometry is not only responsible for the large differences observed experimentally but also provides the principal information required for predicting the resulting rotational interaction.

Conclusion

This study investigated vortex-induced torque during the grasping of compliant objects and quantified the coupled effects of supply pressure, tangential nozzle diameter, surface geometry, and curvature. The results show that increasing the tangential nozzle diameter primarily increases the sensitivity of the generated torque to supply pressure rather than producing a uniformly higher torque. Across all investigated geometries, the median increase in peak torque between 100 and 400 kPa was 23.6%, 53.7%, and 111.6% for nozzle diameters of 0.6, 0.8, and 1.0 mm, respectively. Surface geometry had an equally important influence on the rotational interaction. Dome-concave surfaces generated approximately twice the torque of corresponding dome-convex surfaces in the median matched condition, whereas the difference between cylindrical concave and convex surfaces was substantially smaller. The strongest response occurred for the R=20R = 20 mm dome-concave geometry, which reached 21.770±0.71721.770 \pm0.717 N·mm, with the 1.0 mm nozzles at 400 kPa. A comparison of five machine-learning regression models showed that Gaussian Process Regression provided the best overall predictive performance, achieving R2=0.899±0.080R^2 = 0.899 \pm0.080 under repeated 5-fold cross-validation. The high magnitude and low experimental variation indicate that this response represents a repeatable geometry-dependent phenomenon rather than an isolated extreme value.

Overall, the results demonstrate that vortex-gripper performance cannot be characterized solely by lifting force or supply pressure. This effect should be used for the dexterous rotation of objects being manipulated in both manufacturing and medical applications. Future work will focus on the dexterous manipulation of objects using torque, combining the vortex-induced torque effect.

References

References

  1. [1]G. J. Monkman, Robot grippers. Wiley Online Library, 2007.
  2. [2]J. Zhu, A. Cherubini, C. Dune, D. Navarro-Alarcon, F. Alambeigi, D. Berenson, F. Ficuciello, K. Harada, J. Kober, X. Li, J. Pan, W. Yuan, and M. Gienger, “Challenges and outlook in robotic manipulation of deformable objects,” IEEE Robotics & Automation Magazine, vol. 29, no. 3, pp. 67–77, 2022.arxiv.org/abs/2105.01767
  3. [3]R. Mykhailyshyn, V. Savkiv, P. Maruschak, and I. Xiao, “A systematic review on pneumatic gripping devices for industrial robots,” Transport, vol. 37, no. 3, pp. 201–231, 2022.DOI
  4. [4]G. Papadopoulos, D. Andronas, E. Kampourakis, N. Theodoropoulos, P. S. Kotsaris, and S. Makris, “On deformable object handling: multi-tool end-effector for robotized manipulation and layup of fabrics and composites,” The International Journal of Advanced Manufacturing Technology, vol. 128, no. 3–4, pp. 1675–1687, 2023.DOI
  5. [5]T. Takahashi, K. Nagato, M. Suzuki, and S. Aoyagi, “Flexible vacuum gripper with autonomous switchable valves,” in 2013 IEEE International Conference on Robotics and Automation, 2013, pp. 364–369.DOI
  6. [6]R. Schaffrath, E. Jäger, G. Winkler, J. Doant, and M. Todtermuschke, “Vacuum gripper without central compressed air supply,” Procedia CIRP, vol. 97, pp. 76–80, 2021.DOI
  7. [7]G. Mantriota, “Optimal grasp of vacuum grippers with multiple suction cups,” Mechanism and Machine Theory, vol. 42, no. 1, pp. 18–33, 2007.DOI
  8. [8]J. Y. Park and H. Moon, “Design and control of 2 degree-of-freedom air jet array for manipulating flat objects,” in 2012 9th International Conference on Ubiquitous Robots and Ambient Intelligence (URAI), IEEE, 2012, pp. 467–468.DOI
  9. [9]R. Mykhailyshyn, F. Duchoň, I. Virgala, P. J. Sinčák, and A. Majewicz Fey, “Optimization of outer diameter bernoulli gripper with cylindrical nozzle,” Machines, vol. 11, no. 6, 2023. X. Li, S. Iio, K. Kawashima, and T. Kagawa, “Computational fluid dynamics study of a noncontact handling device using air-swirling flow,” Journal of Engineering Mechanics, vol. 137, no. 6, pp. 400–409, 2011.
  10. [11]R. Mykhailyshyn, V. Savkiv, A. M. Fey, and J. Xiao, “Gripping device for textile materials,” IEEE Transactions on Automation Science and Engineering, vol. 20, no. 4, pp. 2397–2408, 2023.DOI
  11. [12]M. Fujita, S. Ikeda, T. Fujimoto, T. Shimizu, S. Ikemoto, and T. Miyamoto, “Development of universal vacuum gripper for wall-climbing robot,” Advanced Robotics, vol. 32, no. 6, pp. 283–296, 2018.DOI
  12. [13]G. Michalos, K. Dimou­las, K. Mparis, P. Karagiannis, and S. Makris, “A novel pneumatic gripper for in-hand manipulation and feeding of lightweight complex parts—a consumer goods case study,” The International Journal of Advanced Manufacturing Technology, vol. 97, no. 9, pp. 3735–3750, 2018.DOI
  13. [14]S. J. Derby and J. Lippiatt, “Robotic material handling of flexible fuel cell membranes,” in Proceedings of the ASME 2005 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, vol. 47446. American Society of Mechanical Engineers, 2005, pp. 555–563.DOI
  14. [15]C. Tawk, A. Gillett, M. in het Panhuis, G. M. Spinks, and G. Alici, “A 3D-printed omni-purpose soft gripper,” IEEE Transactions on Robotics, vol. 35, no. 5, pp. 1268–1275, 2019.DOI
  15. [16]G. Barrieri-Chávez, S. J. Lázaro-Cárdenas, A. M. Reyes-Duke, and G. J. Amador-Soto, “Adaptive vacuum-suction gripper for efficient object manipulation: Design, simulation, and experimental validation,” in 2026 1th International Conference on Control and Robotics Engineering (ICCRE), 2026, pp. 30–35.DOI
  16. [17]J.-E. Lee, R. Sun, A. Mayur, and L. Sentis, “Grasp failure constraints for fast and reliable pick-and-place using multi-suction-cup grippers,” IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 317–331, 2026.
  17. [19]X. Lyu, K. Shi, and X. Li, “Effect of suction flow in water vortex unit on maximum suction force and flow field,” Physics of Fluids, vol. 37, no. 2, 2025.DOI
  18. [20]K. Shi and X. Li, “Development of a new suction gripper for gripping under-constrained workpiece with minimized contact,” Robotics and Computer-Integrated Manufacturing, vol. 90, p. 102794, 2024.DOI
  19. [21]D. Liu, C. S. Teo, W. Liang, and K. K. Tan, “Soft-acting, noncontact gripping method for ultrathin wafers using distributed bernoulli principle,” IEEE Transactions on Automation Science and Engineering, vol. 16, no. 2, pp. 668–677, 2018.DOI
  20. [22]Y. Zhao, C. Yang, Y. Chen, J. Li, S. Liu, and G. Ye, “Study on the optimal design for cavitation reduction in the vortex suction cup for underwater climbing robot,” Journal of Marine Science and Engineering, vol. 10, no. 1, p. 70, 2022.DOI
  21. [23]C. Wang, J. Zhao, and X. Li, “Effect of chamber diameter of vortex gripper on maximum suction force and flow field,” Advances in Mechanical Engineering, vol. 11, no. 3, p. 168714019837407, 2019.DOI
  22. [24]J. Zhao, C. Wang, and X. Li, “Gap flow with circumferential velocity in annular skirt of vortex gripper,” Precision Engineering, vol. 57, pp. 64–72, 2019.DOI
  23. [25]J.-C. Renn, C.-Y. Chen, and C.-H. Lu, “Gap control for a proportional floating vacuum pad,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, vol. 222, no. 11, pp. 2069–2076, 2008.DOI
  24. [26]B. Ozcelik and F. Erzinçanli, “Examination of the movement of a woven fabric in the horizontal direction using a non-contact end-effector,” The International Journal of Advanced Manufacturing Technology, vol. 25, no. 5, pp. 527–532, 2005.DOI
  25. [27]Y. Wang, W. Zhong, Y. Hu, and X. Gu, “Theoretical and experimental investigation of pressure distribution and suction force of bernoulli gripper,” vol. 2999, no. 1, p. 012031, apr 2025.DOI
  26. [28]R. Mykhailyshyn and J. Xiao, “Influence of inlet parameters on power characteristics of bernoulli gripping devices for industrial robots,” Applied Sciences, vol. 12, no. 14, p. 7074, 2022.DOI
  27. [29]K. Shi and X. Li, “Experimental and theoretical study of dynamic characteristics of bernoulli gripper,” Precision Engineering, vol. 52, pp. 323–331, 2018.DOI
  28. [30]D. Seita, J. Florence, J. Thompson, E. Coumans, V. Sindhwani, K. Goldberg, and A. Zeng, “Learning to rearrange deformable cables, fabrics, and bags with goal-conditioned transporter networks,” in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 4568–4575.arxiv.org/abs/2012.03385
  29. [31]R. Mykhailyshyn, J. Romancik, K. Harada, and A. M. Fey, “Vibration vanquished: Enhancing grasping of deformable objects with jet gripper technology,” in 2025 IEEE 21st International Conference on Automation Science and Engineering (CASE). IEEE, 2025, pp. 2874–2880.DOI
  30. [32]R. Mykhailyshyn, F. Duchoň, M. Mykhailyshyn, and A. Majewicz Fey, “Three-dimensional printing of cylindrical nozzle elements of bernoulli gripping devices for industrial robots,” Robotics, vol. 11, no. 6, p. 140, 2022.DOI
  31. [33]L. Xin, W. Zhong, T. Kagawa, H. Liu, and G. Tao, “Development of a pneumatic sucker for gripping workpieces with rough surface,” IEEE Transactions on Automation Science and Engineering, vol. 13, no. 2, pp. 639–646, 2016.
  32. [34]Z. Sun and Y. Zhang, “Design and research of a high-speed electric vacuum gripper,” Vacuum, vol. 245, p. 114979, 2026.DOI
  33. [35]J. Cao and X. Li, “Noncontact suction mechanism of inhalation-type suction disk,” Physics of Fluids, vol. 37, no. 9, 2025.DOI
  34. [36]T. Alkis, A. M. Fey, and R. Mykhailyshyn, “Automation of studying lifting characteristics of deformable objects for robotic manipulation operations,” in 2026 IEEE/SICE International Symposium on System Integration (SII). IEEE, 2026, pp. 213–218.
  35. [37]Y. Bao, Y. Ambe, and T. Takaki, “Development of a bernoulli suction-type non-contact manipulator capable of controlling object position and orientation during holding,” IEEE Robotics and Automation Letters, vol. 11, no. 6, pp. 7660–7667, 2026.DOI
  36. [38]K. Hanamura, Z. Wang, and S. Hirai, “Fabric manipulation by pulling-driven soft hand with closing-approaching coupling,” in 2024 IEEE/SICE International Symposium on System Integration (SII). IEEE, 2024, pp. 239–244.DOI
  37. [39]P. Zhai, Z. Xu, Y. Yin, X. Li, B. Xie, and H. Wu, “Simulation and experimental analysis of contactless chip pickup process based on a vortex flow gripper,” IEEE Transactions on Semiconductor Manufacturing, vol. 38, no. 2, pp. 324–331, 2025.
  38. [40]D. Liu, M. Wang, N. Fang, M. Cong, and Y. Du, “Design and tests of a non-contact bernoulli gripper for rough-surfaced and fragile objects gripping,” Assembly Automation, 2020.DOI
  39. [41]R. Mykhailyshyn, A. Majewicz Fey, and J. Xiao, “Finite element modeling of grasping porous materials in robotics cells,” Robotica, vol. 41, no. 11, pp. 3485–3500, 2023.DOI
  40. [42]V. Kumar, M. Fontul, C. Neves, and P. J. Coelho, “Prototyping and characterization of gripper technologies for stiff fabric material,” IEEE Access, vol. 13, pp. 107890–107916, 2025.DOI
  41. [43]K. Yamaguchi, Y. Hirata, and K. Kosuge, “Development of robot hand with suction mechanism for robust and dexterous grasping,” in 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2013, pp. 5500–5505.DOI
  42. [44]M. Kojouharov, D. Ho, R. Mykhailyshyn, L. Sentis, and A. M. Fey, “Soft vortex gripper for dexterous manipulation using hand-like robots,” in 2026 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2026, pp. 1–8.DOI
  43. [45]R. Mykhailyshyn, A. Fey, and J. Xiao, “Toward novel grasping of nonrigid materials through robotic end-effector reorientation,” IEEE/ASME Transactions on Mechatronics, vol. 29, no. 4, pp. 2614–2624, 2024.DOI
  44. [46]X. Li, N. Li, G. Tao, H. Liu, and T. Kagawa, “Experimental comparison of bernoulli gripper and vortex gripper,” International Journal of Precision Engineering and Manufacturing, vol. 16, no. 10, pp. 2081–2090, 2015.DOI
  45. [47]R. Mykhailyshyn, V. Savkiv, I. Boyko, E. Prada, and I. Virgala, “Substantiation of parameters of friction elements of bernoulli grippers with a cylindrical nozzle,” International Journal of Manufacturing, Materials, and Mechanical Engineering, vol. 11, no. 2, pp. 17–39, 2021.DOI
  46. [48]J. Zhao, X. Li, and J. Bai, “Experimental study of vortex suction unit-based wall-climbing robot on walls with various surface conditions,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, vol. 232, no. 21, pp. 3977–3991, 2018.
  47. [49]R. Mykhailyshyn and A. M. Fey, “Low-contact grasping of soft tissue using a novel vortex gripper,” in 2024 International Symposium on Medical Robotics (ISMR). IEEE, 2024, pp. 1–6.DOI
  48. [50]R. Mykhailyshyn and A. Majewicz Fey, “Investigating a fluidic vortex grasping mechanism for complex objects and bio-inspired tissue,” IEEE Access, vol. 14, pp. 101032–101048, 2026.DOI

Paper details

vlnc:W_472rxrns
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents