README.md

metadata

license: cc-by-4.0
task_categories:
  - image-to-image
language:
  - en
size_categories:
  - 1K<n<10K

BridgeData V2 Image Triplets Dataset

This dataset contains image triplets from BridgeData V2 trajectories in ImageFolder format.

Derived From

This dataset is a derivative of the 30 GB scripted subset of BridgeData V2 from RAIL-Berkeley. All rights and original licensing apply.

Dataset Structure

• initial_images/: Contains first frame images (initial state)
• intermediate_images/: Contains intermediate frame images (frame 38)
• final_images/: Contains final frame images (frame 43)
• metadata.jsonl: Contains metadata for each trajectory with image file paths

Format

Each trajectory contributes one entry with:

• First image (frame 0) - Initial state
• Intermediate image (frame 38) - Intermediate edited state
• Frame 43 image (frame 43) - Final edited state
• Trajectory metadata (campaign, session, trajectory names, etc.)

All images have a consistent resolution of 640 x 480 pixels.

Statistics

• Total trajectories: 8802
• Total images: 26406
• Image triplets per trajectory: 1
• Image resolution: 640 x 480 pixels

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("imagefolder", data_dir="path/to/dataset")

# Access image triplets
example = dataset[0]
initial_image = example["first_image_file_name"]  # PIL Image - frame 0 (in initial_images/)
intermediate_image = example["intermediate_image_file_name"]  # PIL Image - frame 38 (in intermediate_images/)
final_image = example["frame_43_image_file_name"]  # PIL Image - frame 43 (in final_images/)
metadata = {k: v for k, v in example.items() if not k.endswith("_file_name") and k != "intermediate_image_file_name"}

Citation

If you use this dataset, please cite the original BridgeData V2 paper:

@inproceedings{walke2023bridgedata,
    title={BridgeData V2: A Dataset for Robot Learning at Scale},
    author={Walke, Homer and Black, Kevin and Lee, Abraham and Kim, Moo Jin and Du, Max and Zheng, Chongyi and Zhao, Tony and Hansen-Estruch, Philippe and Vuong, Quan and He, Andre and Myers, Vivek and Fang, Kuan and Finn, Chelsea and Levine, Sergey},
    booktitle={Conference on Robot Learning (CoRL)},
    year={2023}
} 

Generated on: 2025-09-07