GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation
Authors: Haoran Lu, Ruihai Wu, Yitong Li, Sijie Li, Ziyu Zhu, Chuanruo Ning, Yan Zhao, Longzan Luo, Yuanpei Chen, Hao Dong
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate state-of-the-art vision methods, reinforcement learning, and imitation learning approaches on these tasks, highlighting the challenges faced by current algorithms, notably their limited generalization capabilities. Our proposed open-source environments and comprehensive analysis show promising boost to future research in garment manipulation by unlocking the full potential of these methods. |
| Researcher Affiliation | Academia | 1CFCS, School of CS, PKU 2School of EECS, PKU 3Weiyang College, THU |
| Pseudocode | No | The paper describes processes and algorithms but does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | No | We guarantee that we will open-source our code as soon as possible. |
| Open Datasets | Yes | Garment Lab Asset compiles simulation content from a variety of state-of-the-art datasets, integrating individual meshes or URDF files into complete, simulation-ready scenes with robots and sensors. Key components along with their sources and categories are shown in Table 2. More details about each asset type are provided in Appendix A. Clothes Net [70], Shape Net [7], Part Net [35], YCB [72]. |
| Dataset Splits | Yes | We evaluate the generalization ability from the following aspects. Novel Object Thanks to rich Garment Lab Asset, we split garment and other object dataset into Train/Var/Test at proportion 70%/15%/15% to test algorithms generalization ability on object level. |
| Hardware Specification | Yes | Each experiment is conducted on an RTX 3090 GPU, and consumes about 22 GB GPU Memory for training. |
| Software Dependencies | No | For computational resource, we use Py Torch as our Deep Learning framework. Each experiment is conducted on an RTX 3090 GPU, and consumes about 22 GB GPU Memory for training. It takes about 12 hours to train the Coarse Stage, with 1-2 hours of Coarse-to-fine Refinement and 0.5 hour's Few-shot Adaptation. (No version specified for PyTorch). ROS [42] and Move It [10] are mentioned as frameworks/tools, but specific versions used in their implementation are not provided. |
| Experiment Setup | Yes | For Hyper-parameters selection, we set batch size to be 32. In each batch, we sample 32 garment pairs. For each garment pair, we sample 20 positive and 150 negative point pairs for each positive point pair. Therefore, in each batch, 32 * 32 * 20 data will be used to update the model. During the Correspondence training stage, we train the model for 40,000 batches. During Coarse-to-fine Refinement, we train the model for 100 batches. During Few-shot Adaptation, we slightly refine the model using 5 demonstration data. Besides, we set the number of skeleton pairs to be 50. |