RoboParse action segmentation benchmark (v2): test episodes

The list of the 300 test episodes of the RoboParse benchmark, with where each one comes from. Videos and reference annotations are not part of this release; every episode points to its original dataset, where the video can be obtained.

Subset	Episodes	Source
behavior	50	BEHAVIOR-1K 2025 challenge demonstrations, head camera
abc	100	XDOF/ABC-130k, top camera (the benchmark's original ABC test episodes)
abc_extra	50	XDOF/ABC-130k, top camera (episodes added from the former EgoVerse-EVA test set)
egoverse_human	100	EgoVerse, MECKA first-person human demonstrations

Training and validation episodes

The train and validation splits of behavior, abc and egoverse_human list the episodes of the annotator's fine-tuning data (SFT v3, 2026-10-30), in the same format; labels are not part of this release. None of them is a test episode.

Subset	Train	Validation	Notes
behavior	2,000	300	90 tasks, 22-23 episodes each: 45 tasks of the 2025 challenge (task_id 0-49) and 45 tasks added in 2026 (task_id 50-99); task_id gives the task
abc	1,751	88	top camera; includes the 50 former abc_extra test episodes
egoverse_human	800	100	MECKA first-person demonstrations

Since 2026-10-28 the benchmark has 250 test episodes: the 50 abc_extra episodes moved to the abc training split. abc_extra is kept below for reference only. For training rows, duration_sec is the length of the stream the training clips are cut from (ABC-130k: the synchronized 8 fps stream).

Fields

data/<subset>.jsonl, one row per episode:

• episode_id, task (empty for EgoVerse free-form recordings), duration_sec
• source: dataset, and the episode in that release: path (BEHAVIOR, ABC-130k), episode (ABC-130k, EgoVerse), camera, object (EgoVerse bucket object key), egoverse_episode (abc_extra: the EgoVerse copy of the same ABC-130k episode, recorded by the same top camera).

Notes

• ABC-130K is public; models may have seen these episodes during pre-training.
• None of the abc or abc_extra episodes is in the benchmark's ABC fine-tuning data.
• Using the videos is subject to each source's own license and access terms (TODO: link them).

Citation

TODO