Datasets:

JeffrinSam
/

cosmos2.5multip

license: apache-2.0
task_categories:
  - robotics
  - video-generation
  - world-models
tags:
  - cosmos
  - nvidia
  - robot-learning
  - manipulation
  - multi-view
size_categories:
  - n<1K

Cosmos 2.5 Multi-View Robot Manipulation Dataset

This dataset contains multi-view robot manipulation demonstrations formatted for training with NVIDIA Cosmos 2.5 world models.

Dataset Description

Total Episodes: 150
Files per Episode: 4 (3 videos + 1 caption file)
Total Files: 600
Dataset Size: ~1.1 GB
Video Format: MP4
Caption Format: JSONL

Dataset Structure

processeddata/
└── input/
    ├── episode_000001/
    │   ├── caption.jsonl
    │   ├── pinhole_base.mp4
    │   ├── pinhole_side.mp4
    │   └── pinhole_wrist.mp4
    ├── episode_000002/
    │   ├── caption.jsonl
    │   ├── pinhole_base.mp4
    │   ├── pinhole_side.mp4
    │   └── pinhole_wrist.mp4
    ...
    └── episode_000150/
        ├── caption.jsonl
        ├── pinhole_base.mp4
        ├── pinhole_side.mp4
        └── pinhole_wrist.mp4

File Descriptions

Video Files

Each episode contains three synchronized video views:

pinhole_base.mp4: Base/overhead camera view
pinhole_side.mp4: Side camera view
pinhole_wrist.mp4: Wrist-mounted camera view

Caption Files

Each caption.jsonl file contains three lines (one per view) with:

caption: Natural language description of the task
view: Camera view identifier
tag: Additional metadata (nullable)

Example caption.jsonl:

{"caption": "Pick up the bottle and place it into the blue box", "view": "pinhole_base", "tag": null}
{"caption": "Pick up the bottle and place it into the blue box", "view": "pinhole_wrist", "tag": null}
{"caption": "Pick up the bottle and place it into the blue box", "view": "pinhole_side", "tag": null}

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("JeffrinSam/cosmos2.5multip")

# Access episode data
episode_path = "processeddata/input/episode_000001"

Using with Cosmos 2.5

This dataset is formatted for training world models with NVIDIA Cosmos 2.5. Each episode provides:

Multi-view synchronized videos for spatial understanding
Natural language task descriptions
Structured format compatible with Cosmos data loaders

Applications

Robot manipulation learning
Multi-view world model training
Vision-language grounding for robotics
Physical AI simulation
Video prediction models

Citation

If you use this dataset, please cite:

@misc{cosmos2.5multip,
  title={Cosmos 2.5 Multi-View Robot Manipulation Dataset},
  author={JeffrinSam},
  year={2025},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/JeffrinSam/cosmos2.5multip}}
}

License

This dataset is released under the Apache 2.0 License.