molmobot-data

# MolmoBot-data Training episode data (actions, visual inputs, and other sensor data) for 8 tasks on 2 robotic platforms: - DoorOpeningDataGenConfig - RBY1OpenDataGenConfig - RBY1PickDataGenConfig - FrankaPickOmniCamConfig - RBY1PickAndPlaceDataGenConfig - FrankaPickAndPlaceOmniCamConfig - FrankaPickAndPlaceColorOmniCamConfig - FrankaPickAndPlaceNextToOmniCamConfig Please note that every package indexed by the parquet files can contain several instances of episode data. We also provide an extension dataset for `FrankaPickAndPlace` (not used to train any of our [shared models](https://huggingface.co/collections/allenai/molmobot-models)) with additional object types: - FrankaPickAndPlaceOmniCamConfig_ObjectBackfill ## Data access We recommend to use [bulk_download.py](bulk_download.py). For example, ```bash python bulk_download.py --split all --max_part_shards 1 --all /path/to/mbdata/ ``` will download and extract one shard from each part of each configuration in each available split under `/path/to/mbdata`. To launch it you might need to install some dependencies, e.g., by: ```bash pip install zstandard datasets huggingface_hub tqdm ``` A standalone example of how to access the episode data via streaming is also included in [stream_access_example.py](stream_access_example.py). ### Episodes with commercial use-friendly licenses [commercial_episodes-parquet](commercial_episodes.parquet) provides a list with all scenes and episodes where all objects in the scene have a commercial use-frienfly license (aka non-NC) across all data configs, parts, and splits. Example filtering usage: ```python def keep_episode(valid_episodes_string: str, episode_ordinal: int) -> bool: if valid_episodes_string == "*": return True if not valid_episodes_string: return False return str(episode_ordinal) in valid_episodes_string.split(",") ``` `valid_episodes_string == "*"` means all episodes are valid for the current config/split/part/scene (no added objects), else, only the given comma-separated episode indices have all objects under commercial use-friendly licenses. ### Dataset post-processing and stats After bulk downloading data, please use - [repair_video_paths.py](repair_video_paths.py) to fix video paths, - [validate_trajectories.py](validate_trajectories.py) to extract valid trajectories from each dataset and part, and - [calculate_stats.py](calculate_stats.py) to extract dataset stats. For example: ```bash export SPLIT_DIR=/path/to/mbdata/DoorOpeningDataGenConfig/part0/train python scripts/data/repair_video_paths.py "$SPLIT_DIR" python scripts/data/validate_trajectories.py "$SPLIT_DIR" --check-visibility droid_shoulder_light_randomization pickup_obj --check-visibility droid_shoulder_light_randomization place_receptacle python scripts/data/calculate_stats.py "$SPLIT_DIR" --keys actions obs/agent/qpos ``` ## License This dataset is licensed under [ODC-BY 1.0](https://opendatacommons.org/licenses/by/1-0/). It is intended for research and educational use in accordance with Ai2's [Responsible Use Guidelines](https://allenai.org/responsible-use). The subset of Objaverse used to generate these data is licensed under [ODC-BY 1.0](https://opendatacommons.org/licenses/by/1-0/), with license information accessible, e.g., through the [license_info.py](license_info.py) script. The scripts are licensed under [Apache 2.0](CODE_LICENSE). ## BibTeX ``` @misc{deshpande2026molmobot, title={MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation}, author={Abhay Deshpande and Maya Guru and Rose Hendrix and Snehal Jauhri and Ainaz Eftekhar and Rohun Tripathi and Max Argus and Jordi Salvador and Haoquan Fang and Matthew Wallingford and Wilbert Pumacay and Yejin Kim and Quinn Pfeifer and Ying-Chun Lee and Piper Wolters and Omar Rayyan and Mingtong Zhang and Jiafei Duan and Karen Farley and Winson Han and Eli Vanderbilt and Dieter Fox and Ali Farhadi and Georgia Chalvatzaki and Dhruv Shah and Ranjay Krishna}, year={2026}, eprint={2603.16861}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2603.16861}, } ```