metadata
license: apache-2.0
task_categories:
- robotics
- question-answering
tags:
- Navigation
- Retrieval
- Visual Embedding
- Multimodal
- VLM
- MLLM
size_categories:
- n<1K
RAVEN: Long-Horizon Reasoning and Navigation with a Visuo-Spatial-Temporal Memory
Introduction: We release the RAVEN-QA dataset.
Task: given a subsampled list of frames from a long video, each paired with timestamps and positions, and given a user query, the model or robot will find the time or positions where the queried thing shows up. The query can be objects, places, events, and concepts.
Categories: They cover dominant and secondary object retrieval (dominant or secondary in view), visual reasoning, commonsense reasoning, information recall, and spatial understanding.
Diversity: The dataset covers real-world and simulation robot view videos, web-sourced human view videos, and our self-recorded tour videos.
Parts
- irs: A simple retrieval dataset, including YouTube indoor and outdoor videos. (6 videos; 54 queries; with text and image queries)
- irs_hard: A harder human-ego retrieval dataset, including self-recorded and web-sourced videos for more challenging object-finding. (3 videos; 41 queries; with text queries)
- habitat_sim: A robot simulation dataset in Habitat environments. (19 videos; 157 queris; with text queries)
- real_world: A real-world robot exploration dataset taken in our labs and public areas. (4 videos; 21 queries; with text queries)
File Structure
dataset_name/
├─ subsplit_name/
│ ├─ video_name/
│ │ ├─ frames/
│ │ │ ├─ frame_001.jpg
│ │ │ ├─ frame_002.jpg
│ │ │ └─ ...
│ │ └─ questions.json
│ ├─ ...
│ │
├─ subsplit_name/
│ ├─ video_name/
Citation
TODO