Source-linked AI summary
TAPESIM: Efficient Simulation of Adhesive Tape Dispensing for Robotic Manipulation
Zhaofeng Luo, Xinyu Lu, jaehoon Choi, Zhehuan Chen, Trinity Chung, Xiaowen Qiu, Hugh Nicholas Perkins, Gianna Calderon, Alexis Duburcq, Sanghyun Son, Tsun-Hsuan Wang, Yi-Ling Qiao, Minchen Li
TL;DR
Simulating tape manipulation requires capturing a moving roll, a flexible strip, and surfaces that attach and detach, while resolving wound layers can be costly and suppress roll motion. TAPESIM clusters most wound material and advances a deformable collar as tape is released, with optional releasable bonds for interfaces. Clustering improves simulation speed and real-to-sim outcomes in the reported evaluations, while interface rankings vary by task.
Problem
Tape simulation must capture coupled roll motion, payout, and preserved attachment, but resolving every wound layer is costly and can suppress roll motion at practical solver tolerances.
Source: §I · p. 1Method
TAPESIM represents most wound tape as a rigid cluster connected to released tape by an advancing deformable collar, with optional releasable bonds for adhesive interfaces.
Source: §I · p. 1Results
Both clustered variants improve Stick replay error and Peel outcome agreement over full-shell cohesive contact; for Peel, balanced accuracy rises from 50% to 72.9% for Coh-R and 76.3% for Bond-R.
Source: 2) Prediction accuracy · p. 7, §V · p. 8Takeaways & Limitations
The conclusion recommends clustering as the default representation and choosing the interface according to task and cost; teleoperated box sealing demonstrates the complete workflow.
Source: §V · p. 8
Abstract
from arXiv · showhide
Applying adhesive tape to secure wire harnesses or seal packages requires robots to coordinate a flexible strip, a moving roll, and surfaces that attach and detach. Simulation could make these interactions repeatable for robot development and evaluation, but resolving every adhesive layer is expensive and can suppress roll motion at practical solver tolerances, while a permanently rigid roll cannot release material. We present TapeSim, a tape simulator that concentrates deformation near the unwinding region and along the released strip. We will release the source code. A rigid cluster represents most wound material, while an advancing deformable collar enables payout and leaves released tape flexible and reattachable. Optional releasable bonds simplify adhesive interfaces and reduce mean step times for smaller rolls. Controlled swing tests show improved roll rotation. At 32 turns, clustering gives 3.2-3.4x mean physics-step speedups at a fixed Newton tolerance and 4.5-8.4x for comparable roll motion. Across five real-motion Stick replays, the clustered variants reduce mean image-plane core-landmark error by 23-29% relative to the full-shell cohesive baseline. On 100 paired Peel cases, they improve balanced accuracy from 50% to 72.9-76.3%, with interface rankings varying across tasks. A teleoperated box-sealing sequence demonstrates attachment, dispensing, cutting, and sealing in a continuous workflow.
I. INTRODUCTION
Tape simulation must reproduce coupled roll motion, progressive payout, and attachment while avoiding the cost and motion suppression of resolving every wound layer. TAPESIM concentrates deformation near the release front and evaluates simplified adhesive interfaces alongside real-to-sim robot behaviors.
I. INTRODUCTION
- A freely moving tape roll must translate, rotate, and unwind while preserving attachment elsewhere, but fully resolving wound layers is costly and can suppress motion.A permanently rigid roll avoids that burden but prevents payout.Source: §I · p. 1
- TAPESIM connects a rigid cluster of wound tape to the released strip through an advancing deformable collar, keeping dispensed tape flexible and reattachable.This retains deformation near the release front for progressive payout.Source: §I · p. 1
- Optional releasable bonds replace selected contact–adhesion pairs with a single interface energy, and are evaluated for computational benefit and task-dependent prediction accuracy.The paper compares this approach with cohesive contact.Source: §I · p. 1
- The simulator targets attachment, dispensing, and cutting, with evaluation through real-motion replay, frozen-policy outcomes, and continuous teleoperated packaging.The packaging demonstration combines these behaviors in a robot-integrated workflow.Source: §I · p. 1
II. RELATED WORK
- Prior tape systems address specialized tooling or prescribed-attachment placement, whereas TAPESIM models a freely moving roll and progressively released, reattachable tape.Related work also covers adhesive contact and reduced simulation; TAPESIM ties reduction to the payout front and evaluates fixed models through replay and paired policy outcomes.Source: §II · pp. 1–2
III. METHOD
TAPESIM models the tape backing as a thin elastic shell and solves tape and body motion jointly each time step. Its IPC formulation handles contact, friction, and collision prevention.
A. Thin-Shell Simulation with IPC
- The tape backing is a triangular thin shell with StVK membrane elasticity and discrete-shell bending, using the flat rest mesh so winding stores elastic strain.The model describes the nonadhesive strip before adding adhesion.Source: A. Thin-Shell Simulation with IPC · p. 2
- At each time step, the simulator jointly solves tape and body coordinates by minimizing inertial, elastic, normal-contact, and friction terms.The inertial term penalizes departure from inertially predicted motion, while elasticity resists deformation.Source: A. Thin-Shell Simulation with IPC · p. 2
- IPC detects nearby vertex–triangle and edge–edge pairs, applies a barrier as clearance vanishes, and uses continuous collision detection to prevent surface crossing.Projected Newton iterations minimize the objective, with preconditioned conjugate gradients solving the linearized systems.Source: A. Thin-Shell Simulation with IPC · p. 2
B. From Cohesive Contact to Releasable Bonds
TAPESIM offers cohesive contact and releasable bonds for adhesive interfaces. Bonds use a virtual tetrahedron to combine interface responses, while bond creation and release are updated at time-step boundaries.
B. From Cohesive Contact to Releasable Bonds
- The cohesive model tracks each adhesive pair’s engagement intensity; compression promotes engagement, while opening and tangential slip promote debonding.Only the sticky tape face adheres, and a fixed side mask excludes backing-to-backing adhesion.Source: B. From Cohesive Contact to Releasable… · pp. 2–3
- Many wound-layer interfaces barely move, yet their stiff barrier and cohesive terms increase solve cost and can suppress roll rotation.Interface histories and cluster membership remain fixed during each motion solve and update at time-step boundaries.Source: B. From Cohesive Contact to Releasable… · p. 3
- A virtual tetrahedron formed by a contacting point and triangle vertices adds no material or degrees of freedom, and its rigidity energy replaces that pair’s cohesive energy, contact barrier, and friction.Stiffness resists opening, compression, and slip while allowing joint translation and rotation; other contact pairs stay active.Source: B. From Cohesive Contact to Releasable… · p. 3
- A bond forms when an eligible adhesive point–triangle pair enters the bonding distance and releases only in tension when its separation and restoring force exceed the stated criteria.Removing it restores ordinary IPC processing for that pair at the next contact search.Source: B. From Cohesive Contact to Releasable… · p. 3
C. Rigid Wound Region and Advancing Collar
TAPESIM represents most wound tape as a rigid cluster, retaining a deformable collar near the release front and a flexible released strip. The collar advances as tape is dispensed, while the released region can contact and attach to new surfaces.
C. Rigid Wound Region and Advancing Collar
- A rigid cluster represents most wound tape, a deformable collar permits release near the front, and the released region remains a shell for later attachment.The hub remains a separate body coupled to the tape through the tape–hub interface.Source: C. Rigid Wound Region and Advancing… · p. 3
- The cluster shares six translation/rotation degrees of freedom instead of three unknowns per member vertex, omitting member elasticity and internal contacts while retaining boundary and external coupling.The Jacobian transfers vertex-space energy gradients and Hessians into reduced coordinates; collision detection accounts for screw trajectories during line search.Source: C. Rigid Wound Region and Advancing… · p. 3
- A collar of N mesh-graph edges separates the cluster from the free region, with carried positions following cluster motion and collar positions remaining independent shell unknowns.The front update tests deformation away from carried rigid motion and, for bonds, also requires a release event and no remaining live support.Source: C. Rigid Wound Region and Advancing… · p. 3
- As the free region grows, the system restores saved cohesive intensities or original bonds and reference shapes, while released tape never rejoins the original cluster.Cluster formation projects vertex velocities onto rigid motion; vertices becoming deformable inherit the cluster’s instantaneous rigid-body velocity.Source: C. Rigid Wound Region and Advancing… · p. 3
D. Free-Strip Cutting
TAPESIM triggers cutting from blade–tape contact pressure, then removes a finite-width kerf and retriangulates the retained material. The separated pieces continue interacting with the scene.
D. Free-Strip Cutting
- Cutting is triggered when peak local blade–tape contact pressure reaches a prescribed threshold for two consecutive time steps.The pressure uses IPC barrier normal force assigned to a tape element and its reference area.Source: D. Free-Strip Cutting · p. 4
- At a step boundary, the blade curve and nearby triangles are projected onto a local plane, a finite-width kerf is removed, and the retained material is retriangulated.The kerf separates newly exposed edges beyond their combined contact thickness, avoiding coincident surfaces in the next contact step.Source: D. Free-Strip Cutting · p. 4
- New vertices inherit material and kinematic state through parent-triangle interpolation, and shell/contact connectivity and grasp constraints are remapped so both pieces continue interacting with the scene.Source: D. Free-Strip Cutting · p. 4
IV. EVALUATION
The evaluation covers tape handling, computational cost, and real/sim agreement through motion replay and paired policy rollouts, alongside adhesion sensitivity and teleoperated packaging. It uses 19-mm tape assets for experiments except box packaging, with manually verified initial placements.
IV. EVALUATION
- The evaluation measures tape handling, computational cost, and real/sim agreement through motion replay and paired policy rollouts, and also tests adhesion sensitivity and teleoperated packaging.Source: §IV · p. 4
1) Tape Assets:
- Except for box packaging, experiments use 19-mm tape assets whose mass is computed from measured density, flat rest area, and full thickness.The tape is wound as a six-cell-wide triangular shell around a hollow core, with its adhesive face inward; initial placements are manually verified.Source: 1) Tape Assets · p. 4, §IV · p. 4
- Paired real/sim scenes reconstruct initial geometry from the first real frame and include a movable core and the tabletop or target stand.Source: 1) Tape Assets · p. 4
2) Peel-force calibration:
Controlled swing tests show that simplifying wound-roll interfaces and clustering improve rotation, while clustering cuts physics-step time, especially when runs are matched for roll motion. The section also calibrates peel force and demonstrates payout and cutting feasibility.
2) Peel-force calibration:
- The simulated tabletop-peel fixture gives a median reaction of 2.54 N at Cn = Ct = 0.675, against a measured peel force of 2.5 N.The coefficients are fixed for all four Peel models before policy evaluation.Source: 2) Peel-force calibration · p. 4, Figure 2 · p. 5
3) Dispensing and cutting feasibility:
- During 17.75 s of payout, a 32-turn Bond-R roll advances 370.5 mm along the rest-material coordinate as tracked material passes from cluster through collar to free end.A contact-pressure-triggered cut separates the fixed-end strip and releases its tension; the matched no-cut control remains connected through 6.5 s.Source: 3) Dispensing and cutting feasibility · p. 4, Figure 2 · p. 5
B. Roll Motion and Computational Efficiency
- At 32 turns, tightening Coh-F tolerance from 10−2 to 10−3 m/s raises angular range from 9.7° to 38.1°; at 10−2 m/s, Bond-R reaches 37.7° and Coh-R 52.2°.At 32 turns, bonds increase the range to 26.6°; the hub rotation is shown relative to the first post-release state.Source: B. Roll Motion and Computational… · p. 5, Figure 3 · p. 5
- Clustering reduces 32-turn mean step time from 366 to 115 ms within the bond family, a 3.2× speedup; the cohesive family gains 3.4× at the fixed 10−3 m/s tolerance.Timing covers 1750 steps and excludes cluster updates and rendering.Source: B. Roll Motion and Computational… · p. 5
- At 32 turns, clustering reduces mean step time from 307.77 to 36.73 ms in the cohesive family and from 130.43 to 29.14 ms in the bond family, yielding 8.4× and 4.5× speedups for comparable roll motion.The compared runs use tolerances selected to match hub-rotation ranges.Source: Figure 3 · p. 5, B. Roll Motion and Computational… · p. 6
- In a separate 32-turn comparison, Bond-R at 10−3 m/s reduces roll-center RMSE from 11.11 to 5.09 mm and orientation RMSE from 4.68° to 2.17° relative to Bond-F at 10−3 m/s.Bond-F at 10−4 m/s is the numerical reference.Source: B. Roll Motion and Computational… · p. 6
C. Open-Loop Motion Fidelity
Across five real Stick episodes, clustered models better match the roll’s image-plane motion than the full-shell cohesive baseline. Bond-F also shows comparable replay fidelity within the bond family.
C. Open-Loop Motion Fidelity
- The replay metric is Euclidean core-landmark error in undistorted 848 × 480 ego-camera images, averaged across five episodes.Landmarks use the inner-hole ellipse center or, in side views, the outer-core midpoint; timestamps select the nearest simulated state without time-shift fitting.Source: C. Open-Loop Motion Fidelity · p. 6
- 23% and 29% lower mean core-landmark error: Coh-R and Bond-R reach 17.57 and 16.28 px, respectively, versus Coh-F’s 22.96 px.The comparison uses five real Stick episodes replayed with measured arm and finger trajectories.Source: C. Open-Loop Motion Fidelity · p. 6
- Bond-F gives 17.03 px mean error, indicating clustering also retains comparable replay fidelity within the bond family.Errors are averaged after 3 s within each run, across repeats, then equally across the five episodes.Source: C. Open-Loop Motion Fidelity · p. 6
D. Closed-Loop Prediction of Policy Outcomes
Closed-loop evaluation tests whether simulation predicts real-robot outcomes across Peel, Stick, and Wrap. Frozen policies act on paired real and simulated cases, with success rates and case-level prediction measures used to assess agreement.
D. Closed-Loop Prediction of Policy Outcomes
- The three tasks test tabletop peeling, attaching a free tape end to a stand, and dispensing tape around a stand while attaching it.The stand provides a controlled target for Stick and Wrap wire-harnessing subtasks, and each policy acts on its own rollout observations.Source: D. Closed-Loop Prediction of Policy… · p. 6
1) Policies, cases, and metrics:
- Each real test case is aligned to its first image and shared across four simulation configurations, while subsequent robot and tape motions can differ through policy feedback.Initial position, orientation, wound-turn count, and free-end length are randomized.Source: 1) Policies, cases, and metrics · p. 6
- Fine-tuned task policies are selected by validation action R2 and frozen; real and simulated trials use the same weights and preprocessing.Policies receive ego and wrist RGB views, proprioception, and task text; control runs at 100 Hz with 30-Hz waypoints and 3-Hz replanning.Source: 1) Policies, cases, and metrics · pp. 6–7
- Fig. 5 plots simulated success rate against real success rate for 100 Peel, 71 Stick, and 55 Wrap matched cases; the diagonal marks equal rates.Markers distinguish Peel/Stick and Wrap’s one- and two-turn thresholds.Source: Figure 5 · p. 7
- Fig. 6 partitions matched cases into joint success, joint failure, and two disagreement outcomes; balanced accuracy averages recall over real successes and failures.The figure reports paired agreement, while a constant prediction has 50% balanced accuracy.Source: Figure 6 · p. 7, 1) Policies, cases, and metrics · p. 7
2) Prediction accuracy:
Prediction quality varies across tasks: clustering improves individual Peel predictions, clustered models have weak Stick discrimination, and Wrap rankings depend on the success threshold.
2) Prediction accuracy:
- Peel balanced accuracy rises from 50% for Coh-F to 72.9% for Coh-R and 76.3% for Bond-R; agreement increases from 69% to 81% and 82%, respectively.Both clustered variants improve individual predictions over Coh-F.Source: 2) Prediction accuracy · p. 7
- For Stick, Coh-R and Bond-R achieve 67.6% and 69.0% simulated success versus 70.4% real success, but balanced accuracy is only 50.7% and 51.7%.Bond-F makes 48/71 correct predictions and reaches 64.6% balanced accuracy, showing aggregate success-rate agreement does not establish case-by-case prediction.Source: 2) Prediction accuracy · p. 7
- For one-turn Wrap, Coh-R reaches 76.4% agreement and 76.6% balanced accuracy, versus Bond-R’s 72.7% and 72.4%.At the two-turn threshold, Coh-R has 3 false positives versus Bond-R’s 20, while both identify only one of four real successes; Coh-R’s 49/55 agreement includes 48 joint failures, compared with 51/55 agreement from always predicting failure.Source: 2) Prediction accuracy · pp. 7–8
E. Sensitivity to Tabletop Adhesion
TAPESIM’s conclusion recommends clustering as the default representation, with interface choice guided by task and cost. Adhesion sensitivity tests show that stronger tabletop adhesion shifts clustered Peel predictions away from real outcomes, while a teleoperated rollout demonstrates box sealing.
E. Sensitivity to Tabletop Adhesion
- At 20× tabletop adhesion, the sim–real Peel success gap grows from 6.1 to 28.6 percentage points for Bond-R and from 11.2 to 32.7 points for Coh-R; both fail every case at 100×.Coh-F instead reaches 70.4% at 20×, close to the real 69.4%; this coincidence shows aggregate success alone cannot calibrate adhesion.Source: E. Sensitivity to Tabletop Adhesion · p. 8
F. Teleoperated Box Packaging
- In the box-sealing rollout, the cut strip stays attached to the box while the remaining tape stays with the dispenser.The operator closes flaps, anchors and dispenses tape, cuts it, and presses the seal; a 3,846-step rollout advances 128.2 s of simulated time in 304.1 s of wall-clock time (0.422× real-time throughput).Source: F. Teleoperated Box Packaging · p. 8, §V · p. 8
V. CONCLUSION
- At 32 turns, clustering gives 3.2–3.4× mean physics-step speedups at fixed Newton tolerance and 4.5–8.4× for comparable roll motion.The conclusion reports improved Stick replay error and Peel outcome agreement over full-shell cohesive contact; Bond-R provides an additional speed option, while Coh-R performs better on Wrap.Source: §V · p. 8