TeleEmbodied/AssemLM2.0_Datasets

AssemLM 2.0 Processed Assembly Dataset

Related links: AssemLM 1.0 paper (arXiv) · AssemLM official website

This repository provides HDF5 datasets for AssemLM 2.0 training. Each sample contains point clouds for a moving part and a fixed/base part, two assembly manual images, and an object category. The preview/ directory contains a small image-based preview subset configured for display in the Hugging Face Dataset Viewer.

Dataset files

File	Train samples	Test samples	Total	Manual images	Approx. size
assemlm_biasassembly.hdf5	11,653	1,285	12,938	Freestyle	4.33 GiB
assemlm_ikea.hdf5	0	652	652	Freestyle	0.20 GiB
assemlm_partnet.hdf5	51,155	12,897	64,052	Freestyle	8.45 GiB
assemlm_partnext.hdf5	54,743	2,980	57,723	Freestyle	5.59 GiB
assemlm_twobytwo.hdf5	301	140	441	Lineart	0.12 GiB
Total	117,852	17,954	135,806	—	approximately 18.7 GiB

Lightweight subsets for the AssemLM 2.0 GUI

Opening one of the large collections in the AssemLM 2.0 GUI is slow: the dataset panel walks every asset of the selected split before the file becomes usable, which delays both browsing and the inference stage. The sub500/ directory therefore ships smaller versions of the three largest datasets so that the GUI loads them within seconds:

File	Train samples	Test samples	Total	Manual images	Approx. size
sub500/assemlm_biasassembly_sub500.hdf5	500	500	1,000	Freestyle	1.88 GiB
sub500/assemlm_partnet_sub500.hdf5	500	500	1,000	Freestyle	1.88 GiB
sub500/assemlm_partnext_sub500.hdf5	500	500	1,000	Freestyle	1.88 GiB

Each subset keeps the exact structure of its full counterpart — the same split/ and objs/ layout, the same fields, manual image keys, and point-cloud sizes — so the GUI, the dataloader, and the evaluation scripts accept them without any change; they simply contain fewer assets. Every object category of the corresponding full dataset is represented (biasassembly 20, partnet 3, partnext 50), so category browsing and per-category inspection behave exactly as they do on the full files.

These subsets are convenience artifacts for interactive use only: the five full files in the table above remain the reference datasets for training and evaluation.

Hugging Face preview samples

To make the collection inspectable in the Hugging Face Dataset Viewer, the preview/ directory contains a small, separate image-based preview. For every HDF5 file, up to 20 samples are exported from train and up to 20 samples from test.

The current IKEA file has no training samples, so preview/ikea/train contains an empty metadata.jsonl; its preview consists of 20 diverse test samples.

preview/
├── biasassembly/
│   ├── train/
│   │   ├── metadata.jsonl
│   │   ├── 00000_manual_pair.png
│   │   ├── 00000_base.png
│   │   ├── 00000_assemble.png
│   │   └── ...
│   ├── test/
│   └── manifest.json
├── ikea/
├── partnet/
├── partnext/
└── twobytwo/

Each *_manual_pair.png is a side-by-side image containing the base-part and assembled views, so it can be rendered as the primary image column by the Dataset Viewer. The accompanying metadata.jsonl records the asset ID, category, manual type, original image filenames, point-cloud filenames, and point counts. The individual PNG and NumPy files are retained for download and local inspection.

These preview files are additional artifacts only: the five full HDF5 files remain unchanged and continue to provide the complete training/evaluation data.

HDF5 structure

Each HDF5 file contains two top-level groups:

<dataset>.hdf5
├── split/
│   ├── train       # List of training asset IDs
│   └── test        # List of test asset IDs
└── objs/
    └── <asset_name>/
        ├── partA-pc
        ├── base_partB-pc
        ├── category
        ├── image_base_<manual>
        └── image_assemble_<manual>

The <manual> suffix is selected by dataset:

• lineart is used only by assemlm_twobytwo.hdf5;
• The other four datasets use freestyle.

During AssemLM 2.0 inference, the model instruction is constructed from the category:

Assemble the {category} object

Sample fields

Field	Typical shape / type	Description
partA-pc	[N, 3], float32	Moving-part point cloud
base_partB-pc	[N, 3], float32	Fixed/base-part point cloud
category	Scalar string	Object category
image_base_freestyle	[576, 576, 3], uint8	Base-part freestyle rendering
image_assemble_freestyle	[576, 576, 3], uint8	Assembled freestyle rendering
image_base_lineart	[576, 576, 3], uint8	Base-part lineart rendering
image_assemble_lineart	[576, 576, 3], uint8	Assembled lineart rendering

Each sample retains one complete image pair appropriate for its dataset. The point count N is 1,000 or 1,024. The AssemLM dataloader validates the point clouds and, by default, resamples them to 1,024 points.

Loading with h5py

import h5py

path = "assemlm_twobytwo.hdf5"
with h5py.File(path, "r") as f:
    train_ids = f["split/train"][:]
    asset_id = (
        train_ids[0].decode()
        if isinstance(train_ids[0], bytes)
        else str(train_ids[0])
    )
    asset_key = asset_id.replace("/", "_")
    sample = f["objs"][asset_key]

    moving_pc = sample["partA-pc"][:]
    fixed_pc = sample["base_partB-pc"][:]
    category = sample["category"][()]
    base_image = sample["image_base_lineart"][:]
    assemble_image = sample["image_assemble_lineart"][:]

For a freestyle dataset, replace the two lineart keys with image_base_freestyle and image_assemble_freestyle.

Intended use

This dataset is intended for:

• AssemLM 2.0 training, evaluation, and reproducibility studies;
• 3D pose prediction for robotic part assembly;
• Multimodal spatial reasoning from point clouds and rendered images.

Licensing

This release aggregates several upstream 3D assembly datasets, and every subset inherits the terms of the collection it was derived from:

Subset	Upstream dataset	License
assemlm_ikea*.hdf5	IKEA-Manual	CC BY 4.0
assemlm_partnet.hdf5	PartNet	MIT
assemlm_partnext.hdf5	PartNeXt	see the PartNeXt dataset release (its data is built on Objaverse, ABO, and 3D-Future)
assemlm_biasassembly*.hdf5	BiAssemble	see the corresponding dataset release
assemlm_twobytwo.hdf5	Two by Two	MIT

The point clouds and manual renderings in this repository were re-processed and re-rendered for AssemLM 2.0; this release does not supersede any upstream license, and users remain responsible for complying with the terms of the original datasets. If you hold the rights to one of these datasets and would like an attribution adjusted or content removed, please open an issue and we will respond.

Citation

If this project is useful to you, please consider citing:

@article{jing2026assemlm,
  title={AssemLM: A Spatial Reasoning Multimodal Large Language Model for Robotic Assembly},
  author={Jing, Zhi and Qiao, Jinbin and Lu, Ouyang and Ao, Jicong and Qiu, Shuang and Xu, Huazhe and Jiang, Yu-Gang and Bai, Chenjia},
  journal={arXiv preprint arXiv:2604.08983},
  year={2026}
}

Acknowledgements

We thank the authors of the following datasets and research works for their public resources and important contributions:

• IKEA-Manual;
• PartNet;
• PartNeXt;
• BiAssemble;
• Two by Two.

We also thank the authors of Manual-PA and Manual2Skill. Their research on understanding assembly manuals, 3D part assembly, and robotic skill learning provided important references for the construction and application of this dataset.

When using the datasets or research outputs listed above, please follow the licenses and citation requirements of the respective papers, datasets, and code repositories.