| --- |
| pretty_name: WGO-Bench |
| task_categories: |
| - robotics |
| - video-classification |
| language: |
| - en |
| tags: |
| - temporal-segmentation |
| - subtask-annotation |
| - robotics |
| - robot-learning |
| - benchmark |
| license: cc-by-nc-sa-4.0 |
| --- |
| |
| # WGO-Bench |
|
|
| [MacroData's WGO-Bench](https://huggingface.co/datasets/macrodata/WGO-Bench) is a benchmark for evaluating timestamped subtask annotations from robot and egocentric manipulation videos. MacroData describes the benchmark and its annotation approach in [Segmenting Robot Video into Actionable Subtasks](https://macrodata.co/blog/annotating-robot-video-subtasks). |
|
|
| The upstream repository's [`summary.json`](https://huggingface.co/datasets/macrodata/WGO-Bench/blob/9cfb7bb98d60a58319f4b3d7b41374eedfb7980c/summary.json) states that its timeline reannotations were applied only to the 25 `homer_*` episodes, while the DROID and Galaxea rows were left unchanged. During our review of all 100 episodes, we found additional annotation issues. We discuss those issues and our reannotation process in the [VideoDB WGO-Bench reannotation report](https://www.videodb.io/blog/wgo-bench-action-annotations). |
|
|
| This dataset retains the original videos and perception data and replaces the task and timestamped subtask annotations with our reviewed annotations. |
|
|
| ## Dataset |
|
|
| - Episodes: 100 |
| - Corrected subtask segments: 534 |
| - Families: 50 DROID, 25 Galaxea, and 25 HOMER |
| - Format: Parquet |
| - Videos: MP4 bytes embedded in each row |
|
|
| Each row contains the episode ID, video, high-level instruction, timestamped subtask segments, metadata, and available perception state. |
|
|
| ## License |
|
|
| The source dataset is licensed [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/). This dataset uses the same license. |
|
|
