Research map / Reinforcement learning / Visual and representation RL
Learning Efficient Robotic Garment Manipulation with Standardization
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.
Proposes APS-Net, a self-supervised dual-arm policy that selects between dynamic fling and pick-and-place actions to unfold crumpled garments into standardized, aligned configurations, guided by a factorized coverage/IoU/keypoint reward.
Problem
Efficient robotic garment unfolding that also standardizes garment shape and orientation, which simplifies downstream folding, ironing, and packing; prior dual-arm fling methods mainly maximize coverage.
Approach
A unified dual-arm multi-primitive policy (fling vs. pick-and-place) on spatial action maps over rotated/scaled RGB-D inputs, trained with a factorized weighted reward (coverage, IoU, keypoint distance), a spatial action mask restricting actions to reachable garment regions, and an Action Optimized Module that snaps predicted fling points to shoulder keypoints.
Key results
- In simulation on long sleeves, Ours reaches 91.1% coverage, 79.3% IoU, and 1.832 keypoint distance, versus 87.2% / 74.1% / 1.972 for Cloth Funnels.
- In real-world unfolding, Ours achieves 70.5% IoU and 86.3% coverage, above Cloth Funnels (65.2%, 82.2%).
- Real-world folding success is 12/15 for Ours versus 10/15 for Cloth Funnels and 6/15 for FlingBot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use APS-Net instead of Cloth funnels | Long sleeves simulation · Long sleeves simulation | Cov (%): 87.2 -> 91.1 | Helped |
| use APS-Net instead of Cloth funnels | Long sleeves simulation · Long sleeves simulation | IoU (%): 74.1 -> 79.3 | Helped |
| use APS-Net instead of Cloth funnels | Long sleeves simulation · Long sleeves simulation | KD: 1.972 -> 1.832 | Helped |
| use APS-Net instead of Cloth funnels | Long sleeves real-world unfolding · Long sleeves real-world unfolding | IoU (%): 65.2 -> 70.5 | Helped |
| use APS-Net instead of Cloth funnels | Long sleeves real-world unfolding · Long sleeves real-world unfolding | Cov (%): 82.2 -> 86.3 | Helped |
| use APS-Net instead of Cloth funnels | Long sleeves real-world folding · Long sleeves real-world folding | Success Rate: 10/15 -> 12/15 | Helped |
| add KeyPoint Reward (ablation) | APS-Net · Long sleeves simulation | KD: 1.851 -> 1.832 | Helped |
| add Action Optimized Module (ablation) | APS-Net · Long sleeves simulation | Cov (%): 89.5 -> 91.1 | Helped |
| add IoU Reward (ablation) | APS-Net · Long sleeves simulation | IoU (%): 72.2 -> 79.3 | Helped |
| set α=0.2 and β=0.2 instead of α=0.2 and β=0.4 (ablation) | APS-Net · Long sleeves simulation | Cov (%): 91.1 -> 89.9 | Hurt |
| set α=0.4 and β=0.2 instead of α=0.2 and β=0.4 (ablation) | APS-Net · Long sleeves simulation | Cov (%): 91.1 -> 90.7 | Hurt |
| set α=0.4 and β=0.4 instead of α=0.2 and β=0.4 (ablation) | APS-Net · Long sleeves simulation | IoU (%): 79.3 -> 77.5 | Hurt |
Limitations
- Real-world evaluation uses few trials (15 per group) and a single long-sleeve setting for main results.
- Folding relies on heuristics and keypoint detection trained on a limited real-world set (1100 images).
- Reward weights alpha and beta are hand-tuned and the Cloth Funnels comparison is reimplemented in simulation.
Experimental setup
Datasets: CLOTH3D SoftGym simulated garments real-world garment images (1100, LabelMe)
Benchmarks: Cloth Action Gym simulation (long sleeves, pants, skirts) real-world dual-UR5 garment unfolding and folding
Metrics: Coverage IoU keypoint distance success rate
Baselines: p&p S-fling FlingBot Cloth Funnels
Builds on: FlingBot Cloth Funnels SoftGym DeepLabv3
Closest papers by meaning
- UniGarmentManip: A Unified Framework for Category-Level Garment Manipulation via Dense Visual Correspondence (CVPR 2024 · similarity 0.91)
- GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation (NeurIPS 2024 · similarity 0.91)
- DexGarmentLab: Dexterous Garment Manipulation Environment with Generalizable Policy (NeurIPS 2025 · similarity 0.89)
- GarmentPile: Point-Level Visual Affordance Guided Retrieval and Adaptation for Cluttered Garments Manipulation (CVPR 2025 · similarity 0.88)
- Learning 3D Garment Animation from Trajectories of A Piece of Cloth (NeurIPS 2024 · similarity 0.86)
- Reconstruction of Manipulated Garment with Guided Deformation Prior (NeurIPS 2024 · similarity 0.86)
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.