huajian-zeng/gmt-adt-cache

GMT ADT trajectory cache

Preprocessed per-sequence trajectory cache used by GMT (3DV 2026) when training and evaluating on the Aria Digital Twin (ADT) dataset.

Each .pkl file holds the extracted 6-DOF object trajectories, multimodal conditioning features, and indexing metadata for a single ADT sequence, so the GMT training/eval loop does not have to re-parse the raw ADT data on every run.

Paper	arXiv 2603.17993
Code	https://github.com/huajian-zeng/gmt
Pretrained model	huajian-zeng/gmt-adt
Project page	https://huajian-zeng.github.io/projects/gmt/

Contents

• 224 sequence-level .pkl files, named <adt_sequence>_<hash>.pkl
• Total size: ~3.1 GB uncompressed
• Sequence coverage matches the official ADT splits used in the GMT paper

Usage

The simplest path is through the GMT repo's helper script:

git clone https://github.com/huajian-zeng/gmt
cd gmt
bash scripts/download_adt_cache.sh   # pulls this dataset into ./adt_cache/

Or programmatically:

from huggingface_hub import snapshot_download

cache_dir = snapshot_download(
    repo_id="huajian-zeng/gmt-adt-cache",
    repo_type="dataset",
    local_dir="adt_cache",
)

Then point the GMT training / eval scripts at it via --global_cache_dir adt_cache.

License

This release is CC BY-NC 4.0 for research use. The underlying ADT raw data is governed by the Project Aria Research Community License — please review and comply with it.

Citation

@inproceedings{zeng2026gmt,
  title     = {{GMT}: Goal-Conditioned Multimodal Transformer for 6-DOF Object Trajectory Synthesis in 3D Scenes},
  author    = {Zeng, Huajian and Saroha, Abhishek and Cremers, Daniel and Wang, Xi},
  booktitle = {International Conference on 3D Vision (3DV)},
  year      = {2026},
}