Planning a Shared Modular Fixture Layout Across Robotic Disassembly Stages

AI summary

Overview

• Research area: Robotics — robotic disassembly, modular/compliant fixturing, and support-layout optimization.
• Technical level: Advanced (combines quasi-static contact mechanics, force/torque capacity checks, a denoising diffusion probabilistic model, and mixed-variable Bayesian optimization).
• Scope: The paper proposes a framework that plans one shared layout of three vacuum-based balloon hands that stays fixed across every stage of a predefined robotic disassembly sequence for a given product.

What This Paper Is About

As a product is disassembled, each removed component changes the accessible surface, the mass distribution, and the loads the remaining assembly imposes on its fixture, so a support layout that works at one stage may obstruct or inadequately support a later one. Products such as hair dryers, electric drills, and shavers have irregular, asymmetric surfaces that are hard to constrain with rigid fixtures, especially as their geometry changes. The goal is to compute, before disassembly begins, a single set of compliant vacuum-support positions and models that remains feasible under the geometry and task loads of every stage, so the engineer does not reconfigure the fixture between component-removal operations.

Key Contributions

1. Adaptive compliant support system: A modular vacuum-based fixturing system for products with irregular surfaces, built from three balloon hands on independent height-adjustment modules mounted on a magnetic worktable, operated by a KUKA LBR iiwa 14 manipulator with an OnRobot Screwdriver or a Robotiq Hand-E gripper.
2. Cross-stage support-layout optimization: A method that searches for a product-specific shared fixture layout — jointly the positions and models of the three balloon hands — that stays feasible across all predefined disassembly stages, rather than being planned for a single static workpiece state.
3. Physics-guided generative-plus-Bayesian pipeline: A framework that uses a physics-evaluated DDPM to generate support-coordinate seeds and then refines them with Bayesian optimization using a geometry–model mixed kernel for the mixed continuous–discrete configuration space.
4. Qualitative and quantitative validation: Robotic disassembly and directional-load experiments intended to validate the system's support stability (the experimental section's contents are not included in the provided text).

Main Findings

• One layout can be shared across stages: The framework searches a single canonical support configuration that is evaluated at every stage of the sequence, avoiding fixture reconfiguration after each component removal.
• Three balloon hands are adopted: The paper builds on prior work by Pan et al. [1], which applied modular vacuum-based fixturing to screw removal in a single workpiece state and found that three balloon hands provided more stable support than two; the present study adopts that three-hand structure.
• Objects and stage structure: The screwdriver sequence removes the upper housing, motor, and circuit board after the required screw operations; the shaver sequence removes the upper cover, battery, and circuit board. Each object has four states, beginning with the intact product (stages 0–3).
• Four balloon-hand models are considered: SGB10, SGB20, SGB30, and SGB40, with suction coverage radii of 5.0, 10.0, 15.0, and 20.0 mm; physical collision radii of 15.0, 21.5, 33.0, and 37.5 mm; maximum suction forces of 7.5, 23.0, 57.0, and 85.0 N; and torsional resistance limits of 0.06, 0.11, 0.17, and 0.23 N·m, respectively.
• Candidate regions are classified by geometry risk: Surface points are partitioned into a risk region, feasible suction candidates, and a boundary region; risk points are those having a neighbor within a planar radius whose height differs by more than a threshold. Feasible candidates additionally require a local surface-coverage ratio above a threshold. Center mapping uses a 1 mm distance limit and an active-contact coverage requirement of 0.90.
• Contacts can be lost during disassembly: The active set of balloon hands captures contacts that are lost as components are removed; configurations retaining two or three contacts still undergo physical evaluation.
• Failure is modeled by three criteria: Feasibility is determined by finite-contact tipping, tensile suction-capacity violation, and axial torsional-capacity violation. Per-contact demands are estimated by least-squares with minimum-Euclidean-norm allocation, and load components that cannot be represented by these contact actions remain in the residual, which is monitored as a model-consistency diagnostic rather than a hard feasibility criterion.
• Generative seeds concentrate over rounds: The DDPM generates stacked support-coordinate vectors in R^9, and the generated distribution becomes increasingly concentrated from refinement round R0 to R5 under cross-stage physics feedback. The 500 unique records with the highest sequence objective are retained as Bayesian-optimization seeds.
• Bayesian optimization aggregates feasibility across stages: The multi-stage objective combines average and worst-stage feasibility with the robust quality of feasible stages; an infeasible candidate at a stage receives a soft feasibility value that decays exponentially with its violation degree. The final layout is the all-stage-feasible configuration with the highest sequence objective after the budget is exhausted.
• Practical limitations are stated by the authors: The implementation assumes quasi-static operations, predefined object states, and known or estimated task loads; end-effector alignment is performed manually, and unexpected state changes do not yet trigger online fixture replanning.
• Experimental results are not reported in the provided text: The content supplied ends at the beginning of Section IV ("Experimen"), so no experimental numbers, benchmark comparisons, or measured stability margins are available here. The Note to Practitioners states that physical experiments showed the resulting layouts accommodate the tested screw and successive component-removal operations with positive empirical stability margins, but no specific values are given in the available content.

Methodology in Plain English

The researchers define the disassembly task in advance: for each product, four states are specified, starting from the intact object and ending after three component removals, with each state's mesh, mass, center of gravity, candidate support regions, and operation loads known or estimated. The object keeps a nominal pose in which the target screw axes point upward.

Given a candidate layout — three positions plus one balloon-hand model per position — the system checks it at every stage. Each position is snapped in the horizontal plane to the nearest valid suction point on that stage's surface (within 1 mm) and takes the height from that point. Support points that no longer find a valid contact, or whose suction perimeter coverage falls below 0.90, are dropped from the active set. The remaining hands must be far enough apart to avoid collision, given each model's physical radius plus a design gap.

For stability, the plan treats each suction footprint as a circle and builds the region enclosed by the circles' outer common tangents; the object's center of gravity must project inside it with a margin. For each operation, the external forces and torques (gravity plus screw-removal force and torque) are translated to the center of gravity, and the per-contact vertical forces and axial torsional moments needed to balance them are found by least squares. Those demands must stay within the suction force (divided by a safety factor) and the torsional limit of each hand's model. Each stage produces a pass/fail flag and a quality score that blends force demand, risk-region penalty, boundary margin, contact area, and a stability ratio, combined multiplicatively and gated by feasibility.

Searching this space is expensive, so the pipeline is two-stage. First, a denoising diffusion model is trained on physics-evaluated layouts, with the sequence objective as a quality label, and weighted so that higher-quality configurations influence training more. It generates position seeds that are mapped to the stage-0 candidate domain, assigned models, deduplicated, and evaluated; the best 500 become starting points. Second, Bayesian optimization refines the layout. Each candidate becomes a 15-dimensional feature vector (three normalized positions, three normalized boundary margins, and three model indices), and two Gaussian-process surrogates model stage-0 quality and the full-sequence objective, sharing a kernel that multiplies a Matérn-5/2 continuous term by a model term based on differences in suction radius. An acquisition function balances the two objectives and is weighted by a heuristic feasibility score, with a small-probability exploration step. At the end, the archived configurations are re-verified at every stage, and the feasible one with the highest objective is selected.

Why This Matters

• Research impact: The work connects sequence-level disassembly planning with fixture design by making fixture layout a planning variable that must survive changing object states, rather than treating support as given. It also shows a concrete way to inject physics-evaluated generative samples into mixed-variable Bayesian optimization for expensive, stage-coupled objectives.
• Real-world applications:
  • Robotic disassembly cells for small home appliances such as the Panasonic screwdriver and shaver used here.
  • Recycling and material-recovery lines where products must be opened to extract valuable components.
  • Remanufacturing and reuse operations that recover parts from returned or end-of-life products.
  • Any cell handling irregular, asymmetric, or changing part geometry where rigid clamping obstructs access or provides poor support.
• Industry relevance: Preparing one fixture layout before the sequence begins, instead of reconfiguring at each stage, reduces setup interruptions and lets an engineer validate support against the expected operation loads in advance. The paper is explicit that deployment currently needs perception-based alignment, contact monitoring, and online replanning.

Future Directions

• Perception-based end-effector alignment: The current implementation aligns the end effector manually; automated alignment would remove the main manual step.
• Contact monitoring and online replanning: Unexpected changes in the object state do not currently trigger replanning, so sensing-driven updates to the shared layout remain open.
• Relaxing the quasi-static and known-state assumptions: Supporting dynamic operations, unknown or estimated object states, and uncertain task loads would broaden applicability.
• Closing the loop on contact mechanics: The quasi-static model deliberately omits full coupling between deformation, contact area, vacuum pressure, and friction; the paper defers evaluation of model applicability to physical experiments, leaving richer contact models as an open direction.

Target Audience

Researchers and engineers working on robotic disassembly, automated recycling and remanufacturing, and flexible or reconfigurable fixturing; practitioners designing robotic disassembly cells for products with irregular geometry; and method-oriented readers interested in combining diffusion-based generative models with Bayesian optimization for expensive, physics-constrained layout problems.

Authors’ abstract

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.