Research map / Vision-language and multimodal / Robot action and navigation
UniGarmentManip: A Unified Framework for Category-Level Garment Manipulation via Dense Visual Correspondence
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Learns category-level dense point-level visual correspondence for garments via self-supervised contrastive learning (cross-deformation self-play plus skeleton-based cross-object), then adapts it to task-specific functional correspondence with few-shot demonstrations for unfolding, folding, and hanging.
Problem
Garment manipulation (unfolding, folding, hanging) is hard due to diverse garment geometries and deformations; prior methods are task-specific, do not generalize across garments, and rely on large human-annotated data.
Approach
Learn a unified topological dense correspondence across garments of a category and across deformations, using skeleton points for cross-object alignment and self-play point tracing for cross-deformation alignment, then turn it functional with few-shot fine-tuning so one demonstration guides many tasks.
Key results
- Unfolding: Ours achieves 83.6/86.9/81.6% on Unfold-RAND (top/dress/trouser) versus FlingBot 80.3/82.1/79.9%.
- Folding under flinging initial states: Ours achieves 77.9/82.5/81.3% on Fold-FLING, versus ClothFunnels 61.7/63.5/60.3%.
- Real-world folding of 5 tops over 3 deformations each: Ours 11/15 successes versus ClothFunnels 8/15 and UniFolding 10/15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-object correspondence learning (ablation) | UniGarmentManip · Folding (tops, dresses, trousers) | success rate: 75.3 -> 83.6; 70.8 -> 86.9; 77.6 -> 81.6 | Helped |
| add cross-deformation correspondence learning (ablation) | UniGarmentManip · Folding (tops, dresses, trousers) | success rate: 63.9 -> 83.6; 65.2 -> 86.9; 59.4 -> 81.6 | Helped |
| add coarse-to-fine correspondence refinement (ablation) | UniGarmentManip · Folding (tops, dresses, trousers) | success rate: 78.0 -> 83.6; 80.3 -> 86.9; 76.7 -> 81.6 | Helped |
| add few-shot adaptation (ablation) | UniGarmentManip · Folding (tops, dresses, trousers) | success rate: 81.9 -> 83.6; 82.3 -> 86.9; 78.6 -> 81.6 | Helped |
Surprising findings
- On Fold-FLING, the proposed method is slightly below UniFolding on the top category (77.9 vs 78.7), indicating the task-agnostic correspondence does not uniformly beat the task-specific baseline.
Limitations
- Relies on skeleton points and simulation-based point tracing for training labels
- Real-world evaluation is small (5 garments, 3 deformations each) and needs some annotated skeleton points for sim-to-real alignment
- Folding performance on flat garments is comparable to baselines rather than clearly superior
Experimental setup
Datasets: CLOTH3D
Benchmarks: PyFleX simulated unfolding, folding, hanging tasks real-world folding with two Franka Panda arms
Metrics: success rate coverage/IoU success threshold
Baselines: ClothFunnels GCSR DefoAfford FlingBot UniFolding
Builds on: Dense Object Descriptors InfoNCE Skeleton Merger PointNet++ DefoAfford
Closest papers by meaning
- Learning Efficient Robotic Garment Manipulation with Standardization (ICML 2025 · similarity 0.91)
- GarmentPile: Point-Level Visual Affordance Guided Retrieval and Adaptation for Cluttered Garments Manipulation (CVPR 2025 · similarity 0.90)
- GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation (NeurIPS 2024 · similarity 0.90)
- DexGarmentLab: Dexterous Garment Manipulation Environment with Generalizable Policy (NeurIPS 2025 · similarity 0.89)
- DenseMatcher: Learning 3D Semantic Correspondence for Category-Level Manipulation from a Single Demo (ICLR 2025 · similarity 0.86)
- GaPT-DAR: Category-level Garments Pose Tracking via Integrated 2D Deformation and 3D Reconstruction (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.