RiddhiCh/Calvin-Dataset-Final

SGE-Goal CALVIN-ABCD

Training and evaluation data for SGE-Goal, from Goal Images Are Control Targets: Spatially-Grounded Goal-Image Synthesis for Robotic Manipulation (WACV 2027).

• Code: https://github.com/Riddhi-Chatterjee/SGE-Goal
• Project page: https://riddhi-chatterjee.github.io/SGE-Goal/
• Model checkpoints: https://huggingface.co/RiddhiCh/SGE-Goal

Each sample turns one atomic manipulation step from a robot episode into an image-editing example: the scene before the step, the scene after it, where the edit happens, and the instruction. The samples were extracted automatically from CALVIN, ABCD environments, without manual annotation.

Fields

Field	Type	Description
id	string	Sample identifier: the CALVIN episode and atomic segment, e.g. episode_0_atomic_0001
original_image	image	Observation before the manipulation step (I_o)
edited_image	image	Observation after the step, the goal image (I_e)
initial_edit_region_mask	image	Binary mask of the manipulated objects before the step (M_i)
final_edit_region_mask	image	Binary mask of where those objects are after the step (M_f)
edit_instruction	string	The atomic manipulation instruction (T)

Images are 208 × 208 pixels; masks are white (edit region) on black.

The 1,087 test samples are the goal-generation inputs for the closed-loop CALVIN evaluation in the paper (GR-MG and GHIL-Glue); the code repository maps them back to CALVIN episodes.

Splits

Split	Samples
train	22,000
validation	966
test	1,087

Samples are assigned to splits by MD5 hashing of their identifier (seed 42).

Usage

from datasets import load_dataset

ds = load_dataset("RiddhiCh/Calvin-Dataset-Final")
sample = ds["train"][0]
sample["original_image"]      # PIL.Image
sample["edit_instruction"]    # str

How the data was extracted

The SGE-Goal data pipeline chains open vision models on each episode:

1. robot-arm masking and reference-frame selection (EVF-SAM, verified with BLIP-2),
2. class-agnostic object proposals on the reference frame (SAM),
3. object tracking through the episode (XMem),
4. separation of moved and stationary objects with LoFTR matching and 2-means clustering, followed by edit-mask generation and selection of the edited frame.

For episodes without a language instruction, VideoChat-Flash writes one from the video. The pipeline is in data_pipeline/. In a study with 50 annotators rating 500 tuples per dataset (1–5 scale), the lowest mean rating of any extracted component across the three datasets was 4.44.

Intended use

Training and evaluating goal-image generators and language-guided image editing for robotic manipulation. Masks and edited frames come from an automatic pipeline: they are reliable in aggregate, but individual samples can be imperfect, typically with heavily occluded targets or several near-identical objects.

Licence

CALVIN is released under the MIT License; this derived dataset is released under the same licence.

Citation

If you use this dataset, please cite SGE-Goal and the source dataset. The SGE-Goal entry is temporary until the proceedings are published.

@inproceedings{chatterjee2027goalimages,
  title     = {Goal Images Are Control Targets: Spatially-Grounded Goal-Image Synthesis for Robotic Manipulation},
  author    = {Chatterjee, Riddhi and Shrirao, Ritish and Mehrotra, Kushagra and Gopalakrishnan, Viswanath},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}

@article{mees2022calvin,
  author  = {Oier Mees and Lukas Hermann and Erick Rosete-Beas and Wolfram Burgard},
  title   = {CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks},
  journal = {IEEE Robotics and Automation Letters (RA-L)},
  volume  = {7},
  number  = {3},
  pages   = {7327-7334},
  year    = {2022}
}