Soft-Jig Regrasping

• Soft-jig-based regrasping is a robotic manipulation strategy that employs compliant, shape-adaptive fixtures to create stable intermediate states for regrasp operations.
• It integrates soft materials and sensing techniques, such as jamming, pneumatic caging, and tactile feedback, to improve object stability and reduce pose uncertainty.
• The method combines compliance, geometric reconfiguration, and feedback control to enhance grasp precision and enable reliable assembly, regrasp, and disassembly tasks.

Soft-jig-based regrasping denotes a family of robotic manipulation strategies in which a compliant fixture, deformable support, or fixture-like soft contact is used to temporarily stabilize an object while the robot releases, re-approaches, changes grasp pose, or reduces pose uncertainty before a subsequent manipulation or assembly step. In the literature, this role is realized by jamming-based soft fixtures, shell-type pneumatic holders, visuotactile or soft fingertip contacts, and environment-assisted stable placements; across these variants, the common objective is to obtain an intermediate state that is more stable, more repeatable, or more accessible than the original grasp state (Kiyokawa et al., 2020, Kiyokawa et al., 17 Sep 2025, Kolamuri et al., 2021).

1. Conceptual scope and relation to fixture-based regrasping

Soft-jig-based regrasping sits between conventional rigid-fixture regrasping and fixtureless extrinsic manipulation. In classical pick-and-place regrasp, the robot relies on stable intermediate placements and a regrasp graph; adding a simple support pin already expands the set of stable placements and the connectivity of the regrasp graph, because intermediate poses that are impossible on a flat plane become available (Cao et al., 2015). Fixtureless fixturing further abstracts the fixture into an environmental contact: the robot pushes the object against an external contact so that the environment keeps the object stationary while the fingers slide over the object (Chavan-Dafle et al., 2018).

A soft jig preserves the same functional role but changes the mechanics of support. Rather than relying only on rigid support geometry, it uses compliance, friction, and sometimes controllable stiffness to form a shape-adaptive temporary fixture. This is explicit in jamming-based assembly fixtures, where the soft jig acts like a passive hand once the object is placed and jammed, and also in shell-type holders, where inflatable membranes generate a caging condition around the object (Kiyokawa et al., 2020, Kiyokawa et al., 17 Sep 2025).

Related regrasping frameworks broaden the notion of a “fixture” to include the surrounding environment. “Learning to Regrasp by Learning to Place” formulates regrasping from partial point clouds of an object and the supporting environment, using a neural stable placement predictor and a regrasp graph-based solution that leverages and changes the surrounding environment (Cheng et al., 2021). “Multi-Modal Planning on Regrasping for Stable Manipulation” treats sliding, re-grasping, and transferring as sequential modes in an MDP-based planner for rearranging an object into a position suitable for stable manipulation (Hu et al., 2023). These works are not soft-jig systems in the narrow hardware sense, but they make clear that regrasping is fundamentally a problem of generating intermediate support states. This suggests that soft-jig-based regrasping is best understood as a compliant, often reconfigurable instantiation of that broader support-state paradigm.

2. Physical embodiments and contact mechanics

The most direct embodiment is the jamming-based soft jig developed for assembly operations. Its membrane is Dragon Skin FX-Pro silicone rubber with Shore A hardness 2, thickness approximately $1\,\text{mm}$, diameter approximately $160\,\text{mm}$, and bag curvature radius approximately $60\,\text{mm}$; the bag volume is approximately $296\,\text{cm}^3$ and is filled with $450\,\text{g}$ of $\sim 1\,\text{mm}$ glass beads (Kiyokawa et al., 2020). Before fixation, air is pumped in so the beads are free to rearrange and the surface behaves like a soft cushion; during fixation, the part is pressed onto the jig and air is evacuated to confining pressure approximately $90\,\text{kPa}$, producing a jammed, quasi-solid state (Kiyokawa et al., 2020). The resulting contact forces are measured through

$F_n = |f_z|,\qquad F_s = \sqrt{f_x^2 + f_y^2},$

with $F_s \le \mu F_n$ as the effective frictional constraint (Kiyokawa et al., 2020).

A second jamming-derived embodiment is the triangular-pyramid cavity jig for regrasping. Here the fixture is formed by pressing a triangular-pyramid-shaped tool into a silicone membrane bag filled with glass beads, then evacuating the enclosed air so that a stable cavity is created as a placement space (Kiyokawa et al., 17 Sep 2025). The membrane is Dragon Skin FX-Pro, $1\,\text{mm}$ thick, Shore A hardness 2, with internal pressure reduction of roughly $160\,\text{mm}$0; positive pressure is applied before each stamping to reset the membrane to its initial curved state (Kiyokawa et al., 17 Sep 2025). This embodiment shifts the role of the soft jig from direct conformal holding to cavity-mediated pose regularization.

A third embodiment is the shell-type soft jig for holding objects during disassembly. It combines a bottom jamming base with a rigid outer shell containing air chambers lined with a thin silicone membrane. When the bottom jig is depressurized, it provides coarse fixation from below; when the side chambers are pressurized, the membranes inflate inward and create a caging condition around the object (Kiyokawa et al., 17 Sep 2025). The chamber geometry reported in experiments is inner cavity width $160\,\text{mm}$1, shell wall thickness $160\,\text{mm}$2, and chamber height $160\,\text{mm}$3 (Kiyokawa et al., 17 Sep 2025).

A fourth embodiment is not a separate fixture but a soft tactile contact that functionally acts as a local soft jig. GelSight-based regrasping uses a soft elastomer contact surface instrumented by an embedded camera; torque-induced rotation at the grasping point is visualized as rotational marker motion on the elastomer, and this signal is used to drive closed-loop regrasping toward a stable grasp pose (Kolamuri et al., 2021). Closely related work on soft fingertips with Tac3D sensors models the fingertip as a soft contact with a limit surface

$160\,\text{mm}$4

allowing prior grasp perceptions to predict the required grasping force in later postures (Huang et al., 14 Mar 2025).

Embodiment	Primary mechanism	Representative paper
Jamming pad fixture	Shape conformity before vacuum, quasi-solid fixation after jamming	(Kiyokawa et al., 2020)
Triangular-pyramid cavity jig	Tool-stamped cavity for deterministic dropping and regrasp	(Kiyokawa et al., 17 Sep 2025)
Shell-type side jig	Pneumatic caging plus bottom jamming support	(Kiyokawa et al., 17 Sep 2025)
Soft tactile contact	Compliant contact as sensing fixture for closed-loop regrasp	(Kolamuri et al., 2021)

Across these embodiments, two mechanical themes recur. First, compliance broadens the set of physically realizable intermediate placements by conforming to irregular geometry. Second, once fixation is established, stability comes from a combination of normal force, tangential friction, and—where applicable—torsional support from soft contact patches. This suggests that “soft jig” is less a single hardware archetype than a contact-mechanical design pattern.

3. Planning models, configuration algorithms, and grasp-transition structure

The most explicit task-level planning model for a soft jig appears in the assembly-oriented configuration algorithm of “Soft-Jig-Driven Assembly Operations” (Kiyokawa et al., 2020). Assembly parts $160\,\text{mm}$5 are represented through a binary contact matrix $160\,\text{mm}$6 and direction-specific interference-free matrices $160\,\text{mm}$7 for $160\,\text{mm}$8. Reachable directions are then encoded by

$160\,\text{mm}$9

and the reachable-direction list for a pair $60\,\text{mm}$0 is

$60\,\text{mm}$1

Algorithm 1 uses these lists together with center-of-gravity information to choose which part is fixed on the jig and in which posture (Kiyokawa et al., 2020).

When multiple postures are feasible, posture selection is guided by the combined center of gravity

$60\,\text{mm}$2

with preference for lower $60\,\text{mm}$3 (Kiyokawa et al., 2020). In the experiments, low-height, low-CoG postures such as motor P1-a and plate P1-a or P1-c achieved 100% success in fixation under external pushing, while high-height, higher-CoG postures such as motor P2-a and several plate postures achieved 0% success (Kiyokawa et al., 2020). Within the article’s topic, this is a central point: soft-jig regrasping is not merely about compliance, but about compliance configured by geometric reachability and stability analysis.

The triangular-cavity regrasping tool introduces a different optimization problem. For each object and stamping depth $60\,\text{mm}$4, stable placement is evaluated first by a Ferrari–Canny wrench-space criterion,

$60\,\text{mm}$5

accepting the pose as a stable placement pose if $60\,\text{mm}$6; otherwise, a support-polygon fallback is used (Kiyokawa et al., 17 Sep 2025). Grasp feasibility is summarized by $60\,\text{mm}$7, the number of feasible grasp candidates admitting at least one collision-free approach trajectory, while stability margin $60\,\text{mm}$8 is defined either as the distance from gravity wrench to the boundary of the wrench polytope or as the distance from the COM projection to the boundary of the support polygon (Kiyokawa et al., 17 Sep 2025). Depth selection is then posed as

$60\,\text{mm}$9

with $296\,\text{cm}^3$0 in the reported experiments (Kiyokawa et al., 17 Sep 2025).

Regrasp graph formulations in adjacent work provide the broader graph-theoretic context. The support-pin study generates stable placements, computes force-closure grasps, builds a graph of regrasp actions, and searches for regrasp sequences (Cao et al., 2015). “Learning to Regrasp by Learning to Place” similarly centers regrasp around stable placements, taking partial point clouds of the object and supporting environment and outputting a sequence of pick-and-place operations (Cheng et al., 2021). “Synthesizing Grasps and Regrasps for Complex Manipulation Tasks” represents the manipulation plan as a sequence of constant screw motions and determines regrasp necessity from overlaps of graspable regions for contiguous screw segments (Patankar et al., 30 Jan 2025). These formulations imply that a soft jig can be treated as a placement generator, a support-state modifier, or an external contact set within the same general regrasp graph machinery.

A complementary mechanics-oriented viewpoint is given by spring-sliding compliance. In “In-hand Sliding Regrasp with Spring-Sliding Compliance”, each fingertip is modeled as attached to an anchor by a linear spring with contact force

$296\,\text{cm}^3$1

and tangential sliding obeys

$296\,\text{cm}^3$2

under quasi-static wrench balance (Shi et al., 2019). Although that paper focuses on compliant fingertips and a rigid external constraint, its analysis is directly relevant to soft-jig-based regrasping because a soft jig can be modeled as another set of compliant contacts contributing to wrench balance.

4. Sensing, pose inference, and closed-loop regrasp

A defining feature of several soft-jig systems is that the fixture is both support and sensor. “Soft-Jig: A Flexible Sensing Jig for Simultaneously Fixing and Estimating Orientation of Assembly Parts” replaces air-vacuum jamming with a transparent bead-and-oil system, so that two RealSense D435 cameras under the jig can observe markers on the inner membrane surface through refractive-index-matched media (Sakuma et al., 2021). Marker positions are extracted by Laplacian of Gaussian filtering and triangulated into a 3D point cloud; a plane

$296\,\text{cm}^3$3

is then fit by SVD to the deformed membrane region under the object, and the normalized plane normal is used as the principal normal vector of the object (Sakuma et al., 2021). For cylindrical objects with diameter larger than $296\,\text{cm}^3$4, the reported tilt-angle RMSE is less than $296\,\text{cm}^3$5 (Sakuma et al., 2021). This establishes a concrete route from soft-jig deformation to pose estimation, and therefore from fixation to perception-assisted regrasp.

GelSight-based regrasping turns soft contact itself into an incipient-failure detector. The contact markers on the elastomer are tracked, the center of rotation is estimated by a least-squares system $296\,\text{cm}^3$6, and the rotation angle is computed by

$296\,\text{cm}^3$7

A grasp is treated as unstable if the detected rotation exceeds $296\,\text{cm}^3$8, and the robot then shifts the grasp location along the principal axis until rotation falls below the threshold (Kolamuri et al., 2021). The reported dataset contains 142 grasps on 14 objects; the method achieves 94.37% classification accuracy, an onset detection delay of 4.22 frames $296\,\text{cm}^3$9 at 30 fps, and 96.3% success in 109 closed-loop regrasping experiments on 18 objects (Kolamuri et al., 2021).

Multi-sensor slip-based regrasp planning similarly couples tactile sensing to regrasp decisions. “Center-of-Mass-based Robust Grasp Planning for Unknown Objects Using Tactile-Visual Sensors” uses tactile sensors, joint torque sensing, and an RGB-D camera; it reports a slip detection $450\,\text{g}$0 of 76.88% on five unknown test objects and a 31.0% grasp success rate improvement compared to Dex-Net 4.0 (Feng et al., 2020). The central idea is that rotational slip reveals a CoM-related grasp deficiency, and the regrasp planner adjusts grasp pose accordingly. While this system does not use a dedicated soft fixture, it is methodologically adjacent to soft-jig regrasping because the contact interface itself carries the state information needed to choose the next grasp.

The same reuse-of-contact-history theme appears in soft-fingertip force prediction. After one successful grasp, the object gravity vector $450\,\text{g}$1 and center-of-mass position $450\,\text{g}$2 are recovered from measured fingertip forces and moments, then reused to predict the minimum required grasping force in later postures (Huang et al., 14 Mar 2025). Reported prediction errors on controlled objects are at most 0.13 N, 0.18 N, 0.16 N, and 0.12 N across the tested position and rotation cases, with average prediction time approximately 26–29 ms per grasp posture (Huang et al., 14 Mar 2025). In regrasp experiments, the predicted force is used as a feedforward term in a feedback controller, and grasping time in blocked scenarios is reduced from roughly 27–28.3 s for feedback-only control to roughly 5.5–5.7 s for the prediction-assisted controller (Huang et al., 14 Mar 2025). This suggests that a soft jig can serve not only as a stabilizer but as an object-specific memory of contact mechanics.

A more geometric uncertainty-reduction strategy dispenses with fixtures and cameras entirely. “Bimanual Regrasp Planning and Control for Active Reduction of Object Pose Uncertainty” shows that three orthogonal grasps with flat finger pads constrain an object to a unique pose; the deviation levels of the experimental trials are on the same order of magnitude as those of an optical tracking system (Nagahama et al., 28 Mar 2025). Although that method is explicitly fixtureless, it reinforces the broader lesson that regrasp precision can emerge from sequential constraint application, whether the constraints are delivered by a jig, a soft finger, or both.

5. Empirical performance and application domains

The empirical literature spans assembly, disassembly, and general reorientation. In assembly, the jamming pad soft jig was evaluated on a motor and a plate with multiple postures. Shape adaptability was strong at the placement stage: both the motor and the plate remained upright after gripper release in all postures tested, with 10/10 successful trials for all postures of both parts (Kiyokawa et al., 2020). Against external pushing, however, posture mattered sharply. Low-height postures such as motor P1-a and P1-b and plate P1-a and P1-c achieved 100% success, whereas motor P2-a and plate P1-b, P2-a, and P2-b showed 0% success (Kiyokawa et al., 2020). A dual-arm robot then executed pick-and-place of the motor onto the soft jig, insertion of the plate onto the motor, and bolt screwing with an electric driver, with no significant displacement or loss of fixation during insertion and screw-tightening (Kiyokawa et al., 2020). In regrasp terms, the jig successfully acted as a stable third hand throughout a multi-step assembly.

In pose-regularizing regrasp, the triangular-cavity tool emphasizes repeatable dropping before regrasp. Ten mechanical parts were tested, each with 20 drop trials; a trial was successful if position error was at most 5 mm in each axis and orientation error was at most $450\,\text{g}$3 (Kiyokawa et al., 17 Sep 2025). The generated cavity matched the target shape with RMSE at most 4.4 mm, and the reported success rates exceeded 80% for most objects and were above 90% for cylindrical ones (Kiyokawa et al., 17 Sep 2025). Objects (a), (b), and (j) achieved at least 90%, whereas objects (e), (g), (h), and (i) were in the 65–80% range (Kiyokawa et al., 17 Sep 2025). This experimental design is notable because the soft jig is evaluated not by direct grasp retention alone, but by its ability to create a deterministic dropping pose and a stable placement pose suitable for subsequent grasping.

In disassembly and error-tolerant holding, the shell-type jig was compared against a vise and a jamming-gripper-inspired soft jig across ten assembled objects (Kiyokawa et al., 17 Sep 2025). For shaft–bearing, motor–pulley, USB–adapter, AC–adapter, battery–charger, and USB–computer, the shell-type jig remained successful up to $450\,\text{g}$4 of pulling angle deviation, whereas the vise typically tolerated only $450\,\text{g}$5 or $450\,\text{g}$6, and the jamming jig often only up to $450\,\text{g}$7 (Kiyokawa et al., 17 Sep 2025). Force traces during successful pull-out with the shell-type jig stayed around 10 N with low variance, in contrast to the force spikes and rapid drops observed with the vise (Kiyokawa et al., 17 Sep 2025). This is a particularly direct demonstration of a soft jig enabling regrasp-like operations under substantial trajectory error.

Adjacent work on extrinsic manipulation with a rigid support plane provides a useful contrast. A learned placement pipeline using orientation generation, placement refinement, and placement discrimination achieved an accuracy rate of 90.4% and a diversity rate of 81.3% in predicted placements, then used these placements to compute sequential pick-and-place steps for regrasping objects to goal poses not attainable in a single step (Xu et al., 2022). Although the support was rigid rather than soft, the result underscores how much of regrasp performance depends on the quality and diversity of intermediate stable placements. A plausible implication is that soft jigs are most useful when they increase either the stability margin or the diversity of those intermediate states.

Across applications, several recurrent use cases are explicit in the literature: holding a base part during assembly (Kiyokawa et al., 2020), reducing pose uncertainty before regrasp by controlled dropping (Kiyokawa et al., 17 Sep 2025), holding a non-target subassembly during disassembly (Kiyokawa et al., 17 Sep 2025), and correcting unstable grasps by tactilely informed regrasp (Kolamuri et al., 2021). The operational commonality is that the robot relinquishes some direct control to an external support state in order to obtain a better subsequent grasp state.

6. Limitations, misconceptions, and directions of development

A common misconception is that a soft jig is a universal substitute for rigid fixtures. The published results do not support that interpretation. In the jamming pad assembly study, high-height or poorly supported postures failed under external pushing even though all postures were placeable after release; stability required low CoG, sufficient support area, and appropriate datum-plane generation (Kiyokawa et al., 2020). In the triangular-cavity tool, failures were mainly caused by geometric constraints and membrane properties, and even successful placements of some parts were tilted or only marginally stable (Kiyokawa et al., 17 Sep 2025). In the shell-type jig, AC-switch and pulley–shaft showed no success under the tested conditions, indicating that insufficient holding force or unfavorable geometry can defeat a compliant fixture altogether (Kiyokawa et al., 17 Sep 2025).

A second misconception is that compliance automatically improves precision. The opposite limitation is repeatedly reported. In jamming fixtures, repeated deformation and sliding may degrade the silicone membrane, off-board vacuum actuation introduces latency, and deflation/inflation cycles may not reproduce exactly the same contour (Kiyokawa et al., 2020). In the triangular-cavity tool, high elasticity causes bouncing, high friction can prevent sliding down the cavity slope, and filleted stamping-tool edges create smoother cavity corners that weaken positional locking (Kiyokawa et al., 17 Sep 2025). In soft-tactile regrasping, angle estimation degrades at large rotations because adhesion and slip alter the marker-motion model (Kolamuri et al., 2021). In soft-fingertip force prediction, the analytical model assumes local planar contact, uniform friction, and quasi-static conditions; shifting contents, as in the half-bottled juice example, violate the invariance of center-of-mass estimation and reduce success (Huang et al., 14 Mar 2025).

These limitations clarify what soft-jig-based regrasping actually requires. It is not sufficient to provide compliance alone; the system also needs a configuration model, a stability criterion, and a perception or feedback loop. In the literature, those roles are played by reachable directions and CoG-based posture choice (Kiyokawa et al., 2020), by depth optimization balancing stability and grasp accessibility (Kiyokawa et al., 17 Sep 2025), by caging conditions in shell-type jigs (Kiyokawa et al., 17 Sep 2025), by tactile rotation measurement (Kolamuri et al., 2021), or by object-specific force prediction from prior grasp perceptions (Huang et al., 14 Mar 2025).

Future work in the cited literature follows these same axes. The assembly soft jig already points toward pose-estimation integration, automation of jig configuration, and coupling with assembly-sequence optimizers and learning-based assembly strategies (Kiyokawa et al., 2020). The triangular-cavity system explicitly proposes adaptive cavity shaping, automatic tool selection, and integration with pose estimation and assembly sequence optimization (Kiyokawa et al., 17 Sep 2025). GelSight-based rotational regrasping points to multi-fingered robotic arms equipped with tactile sensors (Kolamuri et al., 2021). The soft-fingertip force-prediction framework suggests extensions to richer contact models, dynamic manipulation, and multi-contact soft-jig systems (Huang et al., 14 Mar 2025).

Taken together, the literature indicates that soft-jig-based regrasping is not a single algorithm or device class, but a manipulation regime in which compliant support is made computationally useful through contact modeling, intermediate-state planning, and sensing. Its technical promise lies in reducing dedicated fixture design while preserving enough geometric regularity and force capacity to support reliable grasp transitions. Its present boundary conditions are equally clear: high-CoG or heavily constrained objects, poorly sliding surfaces, shifting mass distributions, and insufficiently modeled soft contact remain difficult cases.