DexGrip: Multi-modal Soft Gripper with Dexterous Grasping and In-hand Manipulation Capacity

In plain words

Picking something up usually means pinching it or wrapping a hand around it. Once it is held, turning it over or sliding it around often means putting it down and grasping it again, which the robot arm has to plan for.

DexGrip is a soft gripper that tries to do that work inside its own hand. Each of its three soft fingers has a belt running over its surface, so the belt can move the object along the finger. A small suction cup on a telescoping arm in the palm can grab an object from above, pull it back, and twist it. Combined, these let the gripper move an object from a fingertip hold to a wider hold and turn it over without setting it down.

The work is a hardware prototype. Its tests cover a set of everyday objects and a few demonstrations, and most results are judged by eye rather than measured.

TL;DR

• DexGrip combines three soft Fin Ray fingers, each with an active surface belt, and an active palm with a telescoping arm and a rotating suction cup §III-A.
• The gripper has 9 degrees of freedom: 3 finger bends, 3 belt rotations and 3 palm motions §III-A.
• The palm test on 12 objects gives 9 overall successes and 10 suction adhesions tab. II.
• In-hand pinch-to-power, a deformable-object reposition and a Rubik's cube face flip are shown, judged qualitatively §IV-C §IV-D.
• Author-reported. No official code was found, and no baseline is measured.

Problem

Grasping and manipulating an object in the hand can cut the arm planning a task needs, but it requires an end effector with more than a grip §I. Soft grippers conform to unknown objects and grasp stably, but their manipulation is limited because the fingers cannot reconfigure while holding a pattern §I §II. Multi-chamber pneumatic fingers add in-hand motion at the cost of a complex pressure system §II. Active surfaces add a degree of freedom, but earlier designs rotate about one axis only §II. The gap the paper names is a design that combines an active palm and active surfaces to get both grasp modes and multi-axis rotation §II tab. I.

Method

The gripper has three subsystems §III-A fig. 1:

1. Fin Ray fingers. Three fingers, each on its own linear actuator, based on the Fin Ray design. The number of ribs was varied and two ribs were chosen for the prototype §III-B §IV-A. Bending gives 3 DoFs.
2. Active surface belt. Each finger has an in-built belt driven by a micro DC motor. It can turn clockwise or counter-clockwise. With enough tension, the object moves with the belt, which gives translation or rotation §III-B §IV-A. The three belts give 3 DoFs.
3. Active palm. A suction cup on a micro DC motor rotates about the x-axis. A Dynamixel-driven telescopic actuator extends and retracts it. The palm sequence is: extend, suction, twist, withdraw §III-B fig. 4. The text counts this as 3 DoFs (extension/retraction, suction, twisting) §III-A.

Suction feasibility is modelled as follows. The lip force is the outer area times atmospheric pressure, minus the inner area times the internal pressure eq. 1. Areas are circles of the lip diameters, A = \pi D^2 / 4 eq. 2. Three loading cases, shown in fig. 2a, give the maximum feasible object mass eq. 3 eq. 4 eq. 5.

Operating sequences tested on the prototype:

• Top-down palm grasp: the palm approaches from above with the cup facing down, then extends, twists and withdraws §IV-B fig. 3.
• Pinch-to-power on a pear: the pear is held at a fingertip, then the belts rotate until its stem points up §IV-C fig. 5b.
• Cube re-orientation with palm and belts §IV-D fig. 5d:

# Cube re-orientation with palm and belts  [§IV-D]
# Input:  cube held by the palm (Rubik's cube, 132 g [tab. II])
# Output: cube with a different face up

def reorient_cube(cube, palm, belts):
    # Stop the palm at the stroke for the desired in-hand position  [§IV-D]
    palm.extend_to_stroke(stroke)
    # Rotate the cube with the suction cup so one flat face points into
    # the gap between two fingers  [§IV-D]
    palm.rotate_cube_into_gap(cube)
    # Engage the palm, then drive all belts in the same direction to lift
    # the cube off the palm  [§IV-D]
    palm.engage()
    belts.drive_same_direction()
    # Rotate the cube about one opposing finger pivot; this raises it  [§IV-D]
    belts.drive_one()
    # Drive all belts back, lowering the cube to its starting height  [§IV-D]
    belts.drive_same_direction_back()
    return cube

Suction feasibility, as written in the paper:

# Lip force and feasible masses  [eq. 1] [eq. 3] [eq. 4] [eq. 5]
# Input:  p_atm, p_suc, d_o, d_i (pressures in Pa, diameters in m)
#         g, mu, a_z, a_h (not stated in the paper)
# Output: f_suc (N), masses m1, m2, m3 (kg)

def lip_force(p_atm, p_suc, d_o, d_i):
    # Outer and inner lip areas, A = pi * D^2 / 4  [eq. 2]
    area_outer = pi * d_o ** 2 / 4
    area_inner = pi * d_i ** 2 / 4
    # Net force as written in the paper  [eq. 1]
    return p_atm * area_outer - p_suc * area_inner

def feasible_masses(f_suc, g, a_z, a_h, mu):
    # Loading case 1, [eq. 3] [fig. 2a]
    mass_1 = f_suc / (g + a_z)
    # Loading case 2, [eq. 4] [fig. 2a]
    mass_2 = f_suc / (g + a_h / mu)
    # Loading case 3, [eq. 5] [fig. 2a]
    mass_3 = mu * f_suc / (g + a_z)
    return mass_1, mass_2, mass_3

Key equations:

F_{\text{suc}} = P_{\text{atm}} \cdot A_o - P_{\text{suc}} \cdot A_i eq. 1

m_1 = F_{\text{suc}} / (g + a_z) eq. 3

m_2 = F_{\text{suc}} / (g + a_h / \mu) eq. 4

m_3 = \mu \cdot F_{\text{suc}} / (g + a_z) eq. 5

Hyperparameters

Most design values are not stated in the paper. Each row says where a value comes from, or "not stated".

Results

Author-reported. Setup: one physical prototype, with a set of 12 objects for the palm test tab. II and a list of test objects with dimensions and weights fig. 3c. No baseline is measured. The degrees-of-freedom comparison uses values from cited papers tab. I.

Paper vs code

No official code. I searched the paper, the arXiv metadata and the LaTeX source for a repository or project page. None is linked. GitHub searches for "DexGrip", "DexGrip soft gripper" and "soft gripper in-hand manipulation" returned no repositories. No code is recorded on this card.

Pitfalls

• Suction sign convention. The text calls the pressure P_{\text{suc}} a "negative vacuum pressure" §III-A. Eq. 1 subtracts P_{\text{suc}} \cdot A_i, which only gives a pulling force if P_{\text{suc}} is the absolute internal pressure. Using −80 kPa directly would add force instead of removing it. Use absolute pressures.
• Loading cases are only in the figure. The three cases in eq. 3 to eq. 5 are shown in fig. 2a and not described in the text. The meaning and sign of a_z and a_h are not stated, so the three masses cannot be reproduced from the text alone.
• Adhesion rate in the text does not match the table. The text says about 75% of objects adhered §IV-B. tab. II shows adhesion for 10 of 12 (83%) and overall success for 9 of 12 (75%). The 75% figure matches overall success, not adhesion.
• Weight box. The text says that once secured, objects "could be re-positioned and re-oriented effectively" §IV-B. In tab. II the weight box (579 g) adheres and can be repositioned, but it fails twisting and overall.
• Mass range. The grasp test is stated to run from 4.6 g to 246.9 g §IV-C. The palm test in tab. II includes a 579 g object, so the two tests do not share a range.
• Aluminium plate. The text says an aluminium plate adhered but collided with fingers and belts §IV-B. The table row for it is commented out in the source, so it is not in the compiled table.
• Degrees-of-freedom count. The abstract and text say the palm has 3 DoFs §I §III-A. The mechanism has a telescope and an x-axis rotation, which are 2 motion DoFs. Suction is not a DoF in the usual sense. Treat the count of 9 as the authors' convention.
• 360-degree rotation claim. The abstract claims full 360-degree rotation about all three principal axes. The palm only rotates about the x-axis, and its twisting "cannot be altered until it interacts with the active surfaces" §IV-B. The experiments show a pear pivot and a cube face flip. No angle range or per-axis measurement is reported.
• Evaluation is qualitative. Grasp stability was judged by moving the end effector and inspecting the object by eye §IV-C. No trial counts, repeats or success rates are given for the in-hand tests, and the palm test is binary per object.
• Stiffness results have no numbers. The FEM section says the results are presented in the experimental section §III-B, but the text gives only qualitative trends §IV-A. Fin Ray material, geometry and the rib counts tested are missing, so the stiffness results cannot be reproduced.
• Belt tension and speed. Tension is described as "appropriate" and no belt speed is given §III-B §IV-D. The conclusion names belt-to-fin friction as a limit §V, so results depend on friction that is not measured.
• Cube sequence. The number of belt cycles, the angles and the timings for the cube flip are not stated §IV-D. The source says the fingers' bending angles were set by hand §IV-D.
• Supplementary video. Several claims refer to the supplementary video: negligible pulling acceleration §IV-B, the aluminium plate collision and the cube flips §IV-D. The video is not part of the arXiv source, so these claims cannot be checked from the text.
• Test objects are fake fruit. The figure caption says fruits are fake artifacts fig. 3. The palm test used a plastic peach §IV-B.
• Equation and section numbers. The equation numbers eq. 1 to eq. 5 were counted from the LaTeX source, where each row of the two-line align environment counts separately. Section numbers use the IEEE roman and letter style. Check both against the PDF before citing.

Lineage

• Builds on: the Fin Ray finger design, Shan and Birglen (2020) §III-B. No card.
• Related active-surface grippers named in the paper: Cai and Yuan (2023), "In-Hand Manipulation in Power Grasp" (the "BACH" entry) §II; Yuan et al. (2020), Roller Grasper V2 §II; Xiang et al. (2024), elastic-band gripper §II. No cards.
• Compared against: six grippers in tab. I, with DoF counts taken from their papers.
• Superseded by: not known.