Source-linked AI summary
TailorCoPilot: Enabling Agentic Pattern Making with Version-Controlled State Tracking
Yuexin Sun, Zhaohui Wang, Ruiyang Liu, Demian Kong, Qian He, Gaofeng He, Huamin Wang
TL;DR
Garment pattern making relies on tacit expertise that existing tools and end-to-end generators do not adequately capture or expose. The paper introduces TailorTrace and TailorCoPilot to record expert revisions as validated, explicit operations and provide state-aware assistance. In a user study, TailorCoPilot improved task success and other reported performance and usability measures, while evidence remains bounded to controlled pattern-revision tasks and held-out traces.
Problem
Garment pattern-making expertise is largely tacit and intermediate drafting steps are poorly captured by existing CAD and generative systems.
Method
TailorTrace represents validated pattern states and semantic transformations, while TailorCoPilot uses recorded traces to provide interactive, editable assistance.
Results
TailorCoPilot significantly increased task-success odds relative to baseline (OR=2.31, 95% CI [1.42, 3.86], p= .001) and improved time, workload, and artifact-quality outcomes.
Takeaways & Limitations
Trace-based support offers a pathway for capturing practice-based expertise and supporting novice and advanced-novice pattern making.
Takeaways & Limitations
Evidence is limited to short-term evaluation in a controlled setting and operation synthesis on held-out traces within a bounded garment domain.
Abstract
from arXiv · showhide
Experience-driven manufacturing, such as garment pattern making, faces a severe generational skills gap because its core expertise relies on undocumented tacit knowledge forged through day-to-day practice. To address this challenge, we present TailorCoPilot, an agentic pattern-making system built upon a specially designed version-control backend TailorTrace. TailorTrace models sewing patterns as structured, discrete states and records their transformations during the pattern-making process as explicit operation sequences defined upon the geometry primitives in the sewing pattern (panels, edges, vertices and stitches). Integrated into a conventional pattern-making GUI, TailorTrace enables seamless documentation of senior experts' tacit pattern-making knowledge without breaking their daily workflow. The documented knowledge further offers interactive, pedagogical scaffolding for novices, while providing a robust foundation to power TailorCoPilot and train future generative AI models. In a user study with novices and advanced novices, TailorCoPilot improved task completion rates, reduced time and perceived workload, and yielded higher-quality artifacts compared to skill-appropriate baselines. Ultimately, TailorCoPilot demonstrates a viable pathway to capture practice-based expertise, operationalizing it to support both generative AI advancements and human apprenticeship.
1 Introduction
Garment pattern making depends on tacit, practice-based expertise that existing computational tools rarely capture. TailorTrace and TailorCoPilot represent pattern development as traceable revisions and use those traces for interactive assistance and apprenticeship.
- Practice-based pattern-making expertise is difficult to formalize because it resides in procedural habits, intermediate decisions, and localized repairs.Finished artifacts alone do not expose how experts preserve fit, balance, and sewability.
- Existing CAD tools provide precision but largely assume users already know the tacit drafting logic required to create patterns.Parametric systems lower learning barriers by constraining designers to predefined structural boundaries.
- Generative pattern-making systems can produce final layouts, but probabilistic outputs may miss industrial tolerances and omit procedural context for inspection or repair.The missing intermediate states make it harder for practitioners to interpret, adjust, or locally repair generated results.
- TailorTrace records pattern states and explicit geometric operations, making expert workflows interpretable, revisitable, and reusable.Its traces provide training data based on expert-defined transformations rather than only static final outcomes.
- TailorCoPilot uses recorded workflows to provide state-aware interactive assistance, with improvements reported in task completion, time efficiency, workload, and artifact quality.The evaluation compared the system with baselines matched to novice and advanced-novice skill levels.
- The paper frames its contributions as process-centered revision, version-controlled validated states, editable agentic support, and evidence that trace-based support improves novice performance.
2 Related Work
Prior garment-design systems support precise geometry, constrained parameterization, or end-to-end generation, but often leave intermediate decisions difficult to inspect and revise. The paper situates TailorCoPilot within broader efforts toward structured, traceable, and interactive AI-assisted design.
- Industrial CAD systems support precise drafting and production workflows but typically assume substantial knowledge of geometric operations, drafting logic, and professional conventions.
- Parametric methods improve efficiency through measurement-driven rules and constrained variables, while reducing designers’ structural freedom.
- GarmentCode and related structured or generative systems encode executable patterns or generate outputs from visual, textual, or multimodal inputs.
- End-to-end garment-generation approaches can produce complete patterns, but their intermediate decisions and revision paths are often difficult to inspect, edit, or reuse.
- Garment pattern making requires context-sensitive decisions because local edits interact with downstream construction constraints and multiple garment pieces.
- Structured computational-design systems represent geometry, constraints, procedures, or rationales, but existing approaches do not fully address manufacturing-oriented revision validity.
- Interactive and agentic AI research emphasizes user agency, control, transparency, iterative interaction, and multi-step reasoning during generation.
3 Formative Study
The formative study examined how pattern makers initiate, execute, evaluate, and repair revisions across expertise levels. Its findings motivate starting from base patterns, capturing expert traces, and keeping revisions inspectable and revisable.
- 3 Formative Study: Pattern making converts a conceptual 3D garment form into precise flat 2D pieces while requiring spatial intuition and attention to physical construction constraints.
- 3.1 Participants and Procedure: The study recruited 13 participants: 6 novices, 4 advanced novices, and 3 experts with more than 5 years of industry experience.
- 3.1 Participants and Procedure: Tasks covered simple local refinements, medium constrained topological operations, and complex reference-driven or structural transformations.
- 3.1 Participants and Procedure: Participants used preferred CAD environments, while retrospective think-aloud protocols, recordings, interviews, and 2D/3D artifacts supported thematic analysis.
- 3.2 Findings: The study found that users often begin from a nearby base pattern and modify it rather than drafting from scratch.
- 3.2.1 Workflow: Experts combine structural principles, prior cases, embodied judgment, and process habits, while less experienced users struggle to access this path-dependent knowledge.
- 3.2.3 Barriers From Intents to Edits: Novices often cannot translate intended appearance into the geometric operation needed to produce it.
- 3.2.3 Barriers From Intents to Edits: Novices rely heavily on visual feedback and may discover downstream sewing or simulation failures only after local 2D changes.
4 System Design
TailorTrace integrates semantic operation tracking, validated pattern states, and unobtrusive version control beneath a conventional CAD interface. TailorCoPilot uses the resulting traces to synthesize editable, state-aware operation sequences from design intent.
- TailorTrace bridges tacit physical-world expertise and a machine-readable version-control backend while preserving a traditional CAD interaction layer.
- 4.1.1 Unified State Representation: Each TailorTrace commit is a validated garment-pattern state containing structured geometry, optional draped geometry, and multimodal metadata.
- 4.1.1 Unified State Representation: The state representation includes 2D panels and stitches, conditionally instantiated 3D draping, and metadata such as descriptions, renders, and fabric parameters.
- 4.1.2 Semantic Action Space and Transformations: Semantic actions such as AddPoint, SpreadPanel, and MovePoint record how and where each pattern evolves between commits.
- Commit-time validation checks 2D closure, intersections, stitching, and—when enabled—3D draping failures before storing states with their full operation sequences.
- Version-control mechanics are embedded behind the CAD GUI so experts can retain familiar mouse or stylus workflows with minimal additional friction.
- 4.2 TailorCoPilot: Agentic Pattern Making: TailorCoPilot translates unstructured design intent into executable operations using structured semantic traces captured from expert workflows.
- 4.2 TailorCoPilot: Agentic Pattern Making: The system uses a dataset of 93 design traces and proposes stepwise, editable operations that progressively transform initialized patterns into intermediate and final states.The traces combine textbook-derived foundational examples with live production traces, while the reasoning module specifies pattern targets, operators, and parameters.
5 Evaluation
The evaluation combined a user study of novice and advanced-novice pattern makers with a diagnostic comparison of trace-augmented reasoning conditions. TailorCoPilot improved task success, completion time, artifact quality, and perceived workload, while expert-authored traces outperformed textbook-only and VLM-only conditions on complex tasks.
- Evaluation Settings: The evaluation measured task success, completion time, expert-rated artifact quality, and NASA-TLX workload in a mixed-design comparison against skill-appropriate baselines.Twenty participants completed randomized tasks across simple, medium, and complex difficulty levels, with blinded expert ratings of sampled artifacts.
- Performance Benefits: TailorCoPilot significantly increased task-success odds relative to baseline (OR= 2.31, 95% CI [1.42, 3.86], p= .001).Novices retained a success advantage across all difficulties, while advanced novices showed the clearest gap on medium and complex tasks.
- Performance Benefits: Completion times were significantly shorter with TailorCoPilot than baseline (p< .001), with the effect size varying by task difficulty (p= .018).TailorCoPilot was generally faster on medium and complex tasks in both expertise groups, although CAD remained competitive for some advanced-novice simple tasks.
- Artifact Quality: Artifact quality was significantly higher with TailorCoPilot (p= .001), with the clearest gains in structural balance and reference conformity.Blinded expert ratings showed consistent but relatively smaller gains in sewability, with strong inter-rater reliability (ICC(1, 1) = 0.81).
- Revision Experience: Perceived workload was significantly reduced with TailorCoPilot according to NASA-TLX scores (p< .001).Exit interviews indicated that participants retained agency and ownership during revision.
- Diagnostic Study: In 15 complex diagnostic tasks, expert-trace augmentation achieved 66.7% success and an average expert rating of M= 4.1 (SD= 0.6), versus 20.0% and M= 1.9 (SD= 0.8) for textbook traces.The difference between textbook-only and expert-trace-augmented conditions was statistically significant (W= 1, p< .001).
6 Discussion
TailorCoPilot makes expert process knowledge usable through traceable, editable revision states rather than final outputs alone. Its benefits extend across novice participation and advanced-novice continuity, while remaining bounded by unresolved expertise, representation, evaluation, and material constraints.
- Reusable Traces as Expert Process Knowledge: TailorCoPilot operates over traceable states, proposes multi-step revisions, and keeps the revision process visible and editable to users.Its agentic behavior depends on a structured revision space that users can inspect and adjust.
- Reusable Traces as Expert Process Knowledge: Expert-authored traces add procedural information for sequencing edits, maintaining cross-piece consistency, and carrying partial results through complex revisions.Textbook-only support produced locally plausible edits but was less reliable at coordinating structurally valid changes across pieces.
- Supporting Apprenticeship Without Replacing Expertise: For novices, TailorCoPilot reduced the CAD fluency and pattern-making knowledge needed to begin meaningful changes.This made complex revision easier to start for novice users.
- Supporting Apprenticeship Without Replacing Expertise: For advanced novices, TailorCoPilot surfaced operations, dependencies, and constraints during revision instead of requiring them to reconstruct expert procedures step by step.This improved continuity through the revision process and made complex revision easier to continue.
- Supporting Apprenticeship Without Replacing Expertise: TailorCoPilot reduced procedural overhead without eliminating the need for expert judgment about balance, silhouette, and construction coherence.Higher-level evaluation and final decisions remained with the user.
- Limitations and Future Directions: The evaluation used a modest sample and controlled short-term setting, while the system remains limited to a predefined operator vocabulary, representable traces, and bounded garment-pattern tasks.The dataset contained 93 traces and approximately 2,500 derived training pairs; broader professional deployment and long-term learning remain unvalidated.
7 Conclusion
TailorCoPilot grounds garment pattern making in TailorTrace, a version-controlled representation of pattern states and transformations. Within the bounded garment-pattern domain, reusable traces improved accessibility for novices and advanced novices and point toward support for apprenticeship and future AI systems.
- Conclusion: TailorCoPilot supports garment pattern making by grounding interaction in TailorTrace, a version-controlled representation of pattern states and transformations.The system uses reusable traces to support revisions involving multiple pieces and intermediate states.
- Conclusion: TailorCoPilot remains bounded by its 2D topological representation, which struggles with intricate physical manipulations such as twisted knots.Figure 9 presents a representative failure case under this constraint.
- Conclusion: Reusable traces made pattern making more accessible to novices and advanced novices, especially for revisions requiring coordination across multiple pieces and intermediate states.The conclusion frames practice-based expertise as more usable when captured as revisitable traces.