oxe-auge/taco_play_inpainting

taco_play robot-removal inpainting dataset

This dataset contains robot-removal inpainting results for taco_play. Each episode provides:

• inpainting.mp4: the robot visually removed via inpainting
• mask.mp4: the robot mask video used for inpainting
• original_episode.mp4: the original (unmodified) episode video
• language_instructions_{split}_all.txt: tab-separated mapping from episode_id to instruction

Relation to OXE-AugE

This release is produced as part of OXE-AugE (AugE-Toolkit), a large-scale robot augmentation project. This inpainting dataset is an intermediate artifact from the overall augmentation pipeline, released independently because it is valuable for downstream research and reuse.

Folder structure

taco_play/
├── README.md
├── archives/
│   ├── taco_play_train.tar
│   └── taco_play_test.tar
├── preview/
│   └── train/
│       ├── 000000_inpainting.mp4
│       ├── 000000_mask.mp4
│       ├── 000000_original.mp4
│       ├── ...
│       └── metadata.jsonl
├── language_instructions_train_all.txt
└── language_instructions_test_all.txt

How to extract

From the dataset repo root:

tar -xf archives/taco_play_train.tar
tar -xf archives/taco_play_test.tar

Each tar extracts to:

taco_play/
└── {split}/
    └── {episode_id}/
        ├── inpainting.mp4
        ├── mask.mp4
        └── original_episode.mp4

Instruction mapping

For each processed split, language_instructions_{split}_all.txt contains lines:

<episode_id>\t<instruction>

(\t means a literal TAB character.)

So episode 17 corresponds to the line starting with 17\t..., and to folder: taco_play/{split}/17/

Citation

@misc{
  ji2025oxeauge,
  title  = {OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning},
  author = {Ji, Guanhua and Polavaram, Harsha and Chen, Lawrence Yunliang and Bajamahal, Sandeep and Ma, Zehan and Adebola, Simeon and Xu, Chenfeng and Goldberg, Ken},
  journal = {arXiv preprint arXiv:2512.13100},
  year = {2025}
}