PaperlayerView exact revision on arXiv

Source-linked AI summary

Planning a Shared Modular Fixture Layout Across Robotic Disassembly Stages

Haohui Pan, Takuya Kiyokawa, Kensuke Harada

arXiv:2608.27151v1cs.RO

TL;DR

Changing geometry, support surfaces, and task loads make single-state fixturing inadequate for complete robotic disassembly sequences. The paper plans a shared modular vacuum layout using physics-informed DDPM initialization and Bayesian optimization, then validates it experimentally. Mean empirical stability margins were 66.9% for the screwdriver and 81.6% for the shaver, and the selected layouts remained applicable throughout the tested sequences.

  • Problem

    Component removal changes support conditions and operation loads, motivating fixture planning over the complete disassembly sequence rather than one workpiece state.

  • Method

    A modular vacuum-based system plans one product-specific shared layout using physics-based cross-stage evaluation, DDPM initialization, and Bayesian optimization.

  • Results

    Mean empirical stability margins were 66.9% for the screwdriver and 81.6% for the shaver.

  • Takeaways & Limitations

    The system determined product-specific shared layouts that remained applicable throughout the tested robotic disassembly sequences.

Abstract

from arXiv · show

Stable support remains challenging in robotic disassembly of irregularly shaped products. As components are progressively removed, the available support surfaces, mass distribution, and task loads change throughout the process. A fixture layout designed for one workpiece state may therefore become infeasible at later stages, motivating unified support planning over the complete disassembly sequence. This paper presents a modular vacuum-based fixturing system that plans one shared support configuration for the complete disassembly sequence of a screwdriver or shaver, allowing each sequence to proceed without fixture reconfiguration. To search the mixed continuous--discrete layout space under repeated cross-stage evaluation, a denoising diffusion probabilistic model generates physics-informed initial configurations that are refined through Bayesian optimization. Robotic screw and component-removal experiments verified the disassembly feasibility of the planned layouts, while 11 directional-load tests quantified their stability. Comparisons between the measured operational loads and directional responses yielded mean empirical stability margins of 66.9% for the screwdriver and 81.6% for the shaver. These results demonstrate that a product-specific shared layout can provide stable support throughout the tested robotic disassembly sequence.

I. INTRODUCTION

Robotic disassembly fixtures must accommodate changing product geometry, support conditions, and operation loads across a known sequence. The paper addresses this need with modular compliant vacuum support and shared-layout planning.

  • Motivation: Component removal changes accessible surfaces, mass distribution, and loading, so a fixture suitable for one state may fail later.
  • Adaptive support: Compliant balloon hands conform to irregular local geometry while vacuum suction retains the object without enclosing it.
  • System: The proposed system integrates three independently height-adjustable balloon hands on a magnetic worktable for screwdriver and shaver disassembly.
  • Contribution: The study extends prior single-state vacuum-fixture evaluation by planning one unchanged, product-specific layout across complete disassembly sequences.
  • Contributions: The paper contributes adaptive compliant support, cross-stage layout optimization under changing geometry and loads, and robotic plus directional-load validation.
  • Related work: Prior work addresses sequence planning, evolving states, operation sensing, fixture construction, reconfiguration, and layout optimization, but generally treats suitable support as given or does not plan one layout across changing states.

C. Generative Design and Bayesian Optimization

The framework searches shared fixture layouts for predefined disassembly stages using physics-guided generative initialization followed by Bayesian optimization. It represents continuous module positions and discrete balloon-hand models while evaluating every stage.

  • Motivation: Existing diffusion and Bayesian-optimization methods do not address support layouts whose feasibility must persist across changing object states.
  • Framework: The framework combines a mathematical support model, DDPM-based initialization, and BO-based layout refinement for expensive mixed-variable search.
  • Candidate representation: Each candidate contains three balloon-hand positions and three discrete model assignments, represented consistently through physical evaluation and BO feature mapping.
  • Optimization loop: The evaluator maps each candidate across all stages, returns stage qualities and a sequence objective, then uses verified outputs to initialize BO with Gaussian-process surrogates.
  • Stage evaluation: The predefined screwdriver and shaver sequences each contain four states, with stage-specific meshes, centers of gravity, masses, candidate regions, and operation loads.
  • Physics model: The quasi-static physics model captures finite-contact tipping, tensile suction capacity, and axial torsional capacity while approximating coupled balloon-hand mechanics.

1) Disassembly Stage and Candidate Representation:

Each disassembly stage supplies state-dependent geometry and loads for mapping a shared support configuration to feasible contact candidates. Surface samples are classified, mapped, and evaluated with available balloon-hand models.

  • Stage inputs: Stage inputs include the object mesh, mass, center of gravity, and operation-dependent force, torque, and application point.
  • Candidate classification: Surface samples are partitioned into risk, feasible-suction, and boundary regions using local geometry and coverage criteria.
  • Candidate classification: Fig. 4 uses red, green, and orange to denote risk, feasible-suction, and boundary regions, respectively.
  • Stage mapping: Each support center is mapped in XY to the nearest feasible candidate within 1 mm, while its height comes from the stage candidate.
  • Active contacts: The active contact set captures contacts lost after component removal, and configurations with two or three contacts proceed to physical evaluation.
  • Hand models: Four balloon-hand models provide lookup values for suction coverage radius, physical collision radius, maximum suction force, and torsional resistance.

2) Geometric Feasibility Constraints:

Geometric and load feasibility constrain shared layouts against contact spacing, boundaries, finite-contact support, and operation-specific force–torque demands. The model evaluates modeled physical failure criteria across stages.

  • Geometric constraints: Geometric feasibility combines active-contact testing, inter-support collision avoidance, and boundary margin.
  • Stability: Finite-contact stability requires the center of gravity projection to lie inside the support region with the configured minimum tipping margin.
  • Operation loads: Stage loads are specified by task force, direct task torque, and application point under the prescribed screw-operation pose.
  • Operation loads: Stages without screw removal use zero task force and torque, producing the gravity-only case.
  • Load allocation: The wrench mapping distributes vertical contact forces and axial torsional moments among active contacts using their positions relative to the center of gravity.
  • Load allocation: When multiple least-squares allocations exist, the minimum-Euclidean-norm solution provides deterministic per-contact demands.
  • Feasibility criteria: Residual loads are monitored diagnostically, while feasibility uses finite-contact tipping, tensile suction-capacity, and axial torsional-capacity failure criteria.

4) Support Quality Score:

The support-quality score combines physical penalties, geometric rewards, and operation-pass performance, while hard feasibility gates the result. Candidates failing required checks receive zero quality.

  • Each stage returns a hard-feasibility indicator hk,s and a quality score Qk,s in [0, 1].The same score supports DDPM feedback, Bayesian optimization, and archive verification.
  • The score combines normalized force-demand and risk penalties, boundary-margin and contact-area rewards, and operation-pass ratio.All component terms are normalized to [0, 1].
  • Component weights and shaping parameters control the multiplicative combination of score terms.The combined score is gated by hard feasibility.
  • A candidate is feasible only when mapping, coverage, collision, nominal gravity stability, and all required operation-load checks pass.CoG-uncertainty cases contribute continuously through Sstab,k,s; infeasible candidates receive Qk,s = 0.

C. DDPM-Based Initialization

The DDPM generates physics-informed support-coordinate seeds for complete-sequence optimization, then passes selected layouts to Bayesian optimization. Its generated distributions become more concentrated under cross-stage feedback.

  • The DDPM learns from physics-evaluated configurations because many canonical-state-feasible samples fail later in the disassembly sequence.It generates seeds with higher complete-sequence quality for BO initialization.
  • The model operates on a stacked nine-dimensional support-coordinate vector X(k), while balloon-hand models are assigned during physics-based postprocessing.The sequence objective Jk serves as the physics-based quality label.
  • The network predicts sampled Gaussian noise and the sequence objective using a timestep-conditioned multilayer perceptron with two output heads.Training combines noise-prediction and objective-prediction losses.
  • Quality-dependent loss weights increase the influence of higher-quality configurations during DDPM training.The implementation uses wϵ(Jk) = 0.01 + 0.99Jk and wJ(Jk) = 0.2 + 0.8Jk.
  • At refinement rounds R0–R5, generated support-coordinate distributions become increasingly concentrated under cross-stage physics feedback.The 500 unique records with highest Jk are retained as BO seeds.
  • Bayesian optimization refines the seeds because the cross-stage objective is expensive, gradient-free, and repeatedly evaluated across changing stages.Its candidate features include normalized support coordinates, boundary-margin features, and balloon-hand model indices.

2) Geometry–Model Mixed-Kernel Surrogates:

The BO surrogates represent both continuous support geometry and discrete hand-model choices, while the multi-stage objective aggregates feasibility and quality across the sequence. Selection balances expected improvement for canonical quality and sequence performance.

  • Two Gaussian-process surrogates model stage-0 physical quality Qk,0 and complete-sequence objective Jk.The first guides canonical quality improvement; the second represents sequence feasibility and robustness.
  • A mixed kernel captures similarity in continuous geometry and balloon-hand model-dependent suction-radius differences.The kernel combines normalized positions, margins, and model features.
  • Each infeasible stage receives a violation degree based on geometry failure, stability shortfall, failed load cases, and boundary-margin deficiency.The four terms are weighted 1, 0.8, 0.8, and 0.5, respectively.
  • Soft feasibility equals one for feasible stages and exp(−βvk,s) otherwise, with β controlling the infeasibility penalty.This converts graded violations into a continuous stage-level feasibility value.
  • Cross-stage feasibility combines average and worst-stage feasibility over the complete disassembly sequence.The weights wmean,C and wmin,C sum to one.
  • Robust quality combines worst- and mean-stage quality over feasible stages, then the multi-stage quality is gated by cross-stage feasibility.If no feasible stage exists, Qrob,k is set to zero.
  • Candidate selection uses expected improvements for stage-0 quality and the sequence objective over the unevaluated candidate set.The objective parameters are optimization settings rather than fixed physical parameters.

4) Acquisition-Based Candidate Selection:

Acquisition-based selection favors candidates with estimated physical feasibility, supplements acquisition ranking with limited exploration, and evaluates each selected candidate across all stages. The final archive selects the highest-Jk feasible layout.

  • The feasibility heuristic combines candidate-region validity, geometric feasibility, and minimum normalized stage-0 boundary margin.Its implementation is gfeas_t,k = 0.3 + 0.7fproxy,t,k.
  • A small-probability exploration step may select from an acquisition-ranked subset or the remaining candidate pool.This supplements the primary acquisition-based ranking.
  • Each selected candidate is evaluated at all stages, and its quality updates both surrogate models.This preserves the cross-stage evaluation used by the objective.
  • After the evaluation budget is exhausted, deduplicated archive configurations are verified across all stages and the feasible candidate with highest Jk is selected.Figure 6 reports cumulative unique all-stage-feasible configurations and explored locations with selected layouts.

IV. EXPERIMENTS

The planned shared layouts were tested against vise limitations, robotic disassembly tasks, screw operations, and directional-load measurements. Experiments evaluated practical feasibility, operational responses, and resistance under changing disassembly conditions.

  • Fixture comparison: A parallel-jaw vise either lost fixation on the screwdriver or obstructed shaver access under the required screw-access pose.The screwdriver’s curved housing and asymmetric mass caused tilting, while the shaver’s cover and gripper clearance were obstructed.
  • Screw operations: Screw removal and tightening were completed at all seven locations without changing the shared fixture layout.Removal used an 18 N axial load, while tightening applied a 0.5 N m torque.
  • Screw operations: 86.4% of 140 screw-removal attempts succeeded overall, with location-wise rates ranging from 60% to 100%.Observed failures were primarily associated with screwdriver-bit and screw misalignment, outside the fixture-layout scope.
  • Disassembly feasibility: All 60 component-removal trials succeeded without changing the planned shared fixture layout across the tested sequences.Each object underwent three sequential removals using independent repositioning trials.
  • Operational loads: Operational force and moment peaks varied across objects and stages, reflecting changes in geometry, support pose, removed components, and contact locations.The largest force and moment responses did not necessarily occur at the same disassembly stage.

C. Directional External-Load Resistance

Directional resistance was measured across 11 commanded translational and rotational directions, with repeated repositioning trials and incremental force–moment analysis. The responses revealed object-dependent anisotropy and bounded the interpretation of measured support resistance.

  • Test design: Directional resistance was measured in five translational and six rotational commanded directions, with five independently repositioned trials per direction.The direction labels refer to robot end-effector commands; A and B identify the screwdriver and shaver.
  • Test design: The z-n direction was excluded because downward loading compresses the support rather than testing loss of support, while z-p tests upward separation.Higher z-n loading could damage the linear modules without producing meaningful detachment.
  • Response analysis: The tests reported coupled incremental force–moment responses because commanded motion and load application points need not align with sensor axes.Component values therefore preserve measured coupling rather than acting as independent limits.
  • Measured resistance: Screwdriver translational and rotational resultants ranged from 27.62 N to 56.13 N and 3.79 N m to 5.60 N m, respectively.The corresponding shaver ranges were 32.49 N to 88.75 N and 1.51 N m to 6.38 N m.
  • Measured resistance: Both objects showed relatively large x-n translational resistance but different rotational anisotropy linked to geometry, mass distribution, support configuration, and load location.The screwdriver’s offset mass and L-shaped geometry affected moment arms, while the shaver showed lower response about a than about b or c.

V. COMPUTATIONAL EVALUATION AND DISCUSSION

The paper converts directional responses and operational loads into empirical stability margins under an unfavorable-direction assumption. These margins remain positive for the tested layouts and average 66.9% for the screwdriver and 81.6% for the shaver.

  • Margin formulation: The unfavorable-direction assumption compares each complete operational force or moment resultant separately with every corresponding directional response.It does not imply simultaneous loading in all directions.
  • Margin formulation: A larger ξd,o denotes a larger measured directional resistance margin relative to the operational load.The operation-level mean averages the 11 measured directions equally.
  • Margin interpretation: The direction-wise and mean margins quantify measured stability reserves for each shared layout under the tested operations and loading conditions.The indicator retains direction-wise values to expose anisotropic behavior.
  • Results: Screwdriver operation-level margins were 86.4%, 76.4%, and 86.0% for dis-1, dis-2, and dis-3, respectively.The corresponding shaver margins were 78.9%, 84.1%, and 81.6%.
  • Results: 66.9% and 81.6% were the object-level mean empirical stability margins for the screwdriver and shaver, respectively.All direction-wise margins remained positive under the defined assumption.
  • Scope: The reported margins depend on the tested layouts, object geometry and mass distribution, contact states, load-application locations, and unfavorable-loading assumption.They are scoped to the tested disassembly operations and loading conditions.

B. Multi-Stage Bayesian Optimization Ablation

The ablation evaluates multi-stage Bayesian optimization settings and DDPM initialization for finding layouts feasible across changing disassembly stages. The selected setting and DDPM seeds improved cross-stage feasibility, followed by physical validation of the resulting layouts.

  • Objective-setting ablation: 57.7% all-stage-feasible rate made J4 the selected objective setting, despite J10 achieving higher later-stage trimmed-mean qualities.J4 also achieved a mean feasible-stage ratio of 0.831 and mean violation degree of 0.100.
  • DDPM-seed ablation: DDPM initialization produced 577 all-stage-feasible evaluations and 79 unique feasible layouts, whereas structured-random initialization produced none.The comparison used 500 initial candidates and 1000 subsequent Bayesian-optimization evaluations under identical settings.
  • DDPM-seed ablation: DDPM initialization increased the mean feasible-stage ratio from 0.500 to 0.831 and raised the trimmed-mean objective from 0.348 to 0.711.It also reduced the mean violation degree from 0.386 to 0.100 in the controlled single-seed comparison.
  • Physical validation: The planned shared layouts supported robotic screw operations and successive component-removal experiments without fixture reconfiguration.The layouts were evaluated physically on screwdriver and shaver cases.
  • Physical validation: 66.9% and 81.6% were the mean empirical stability margins for the screwdriver and shaver, respectively, under directional-load testing.These margins compared measured directional responses with operational loads.
Loading 2608.27151v1…