File size: 1,322 Bytes
60baccb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
---
license: mit
task_categories:
- video-text-to-text
- visual-question-answering
tags:
- robotics
- robomme
- manipulation
- memory
- video-qa
size_categories:
- n<1K
configs:
- config_name: episodes
  data_files: episodes.parquet
---

# RoboMME — SequenceRecoveryVertically (Video QA)

Video-QA dataset for the **SequenceRecoveryVertically** task from
[RoboMME](https://github.com/), a ManiSkill/SAPIEN benchmark for
memory-augmented robotic manipulation. The agent watches a demonstration video,
remembers the arrangement of cubes, and rebuilds it around a pre-placed anchor
cube before pressing a stop button.

## Contents

- `episodes.parquet` — 500 train episodes with per-episode metadata (seeds,
  difficulty, task semantics, language goals, success flags, video paths).
- `videos/` — 1000 MP4 clips: one execution video and one `_recall` demonstration
  clip per episode.
- `*_mc_qa_processed.json` — multiple-choice QA pairs (LLaVA conversation format).
- `*_oe_qa_processed.json` — open-ended QA pairs.
- `*_cap_processed.json` — caption pairs.
- `collection_manifest.jsonl` — provenance manifest (source H5, SHA-256 hashes,
  frame counts).

QA/caption entries reference clips via a `video` field (e.g.
`videos/..._recall.mp4`) and use the standard `conversations` schema with
`<image>` tokens.