language:
- en
license: cc-by-4.0
size_categories:
- n<1K
task_categories:
- robotics
- reinforcement-learning
pretty_name: RoboCombine Dataset
tags:
- robot-manipulation
- compositional-generalization
- vision-language-action
- benchmark
- simulation
configs:
- config_name: default
data_files:
- split: test
path: data/RoboCombine_episodes.parquet
RoboCombine Dataset
Paper: Exploring the Limits of Combinatorial Generalization in Embodied Manipulation
Code: https://github.com/AutoLab-SAI-SJTU/RoboCombine
RoboCombine is a compositional robot-manipulation dataset release. The Dataset
Viewer split is backed by data/RoboCombine_episodes.parquet, an episode-level
manifest with 950 rows for 19 tasks x 50 episodes.
The release archives remain available for reproducibility:
RoboCombine_eval_repo-0.1.0.tar.gz: evaluation repository, task specification, metadata, and runner code.RoboCombine_data-0.1.0.tar.gz: optional HDF5 trajectories, scene metadata, seeds, and MP4 previews.
The manifest uses task identifiers, instructions, relative artifact paths, and frame counts. It does not contain private logs, model checkpoints, or local filesystem paths.
Scale summary:
- 19 tasks
- 50 episodes per task
- 950 total episodes
- 285806 total recorded frames
Croissant metadata is available through the Hugging Face /croissant endpoint after the
Dataset Viewer processes the Parquet manifest. A repository-local copy is also provided at
metadata/croissant.jsonld.
Repository id: AutoLab-SJTU/RoboCombine.
License
The dataset contents in this repository, including the episode manifest,
metadata, and release archives, are released under CC-BY-4.0. Code shipped
inside the evaluation archive retains its own bundled MIT license notice.
Third-party assets, simulators, checkpoints, and interfaces referenced by the
release archives remain subject to their original upstream licenses.