Apache 2.0
The publisher's licence permits commercial use and redistribution, with attribution. Hub card tag apache-2.0, repeated as License: apache-2.0 in the card body; no longer licence text in the repo.
hf download allenai/MolmoAct2-BimanualYAM-Dataset --repo-type dataset --include meta/* --local-dir ./MolmoAct2-BimanualYAMPublic, no gate. About 2.35 TB; the command pulls only the metadata and task annotations. AV1 video. The other repos in the MolmoAct2 collection are LeRobot conversions of DROID and Open X-Embodiment parts, community SO-100 data or simulation, and are not this row.
Specification, as published
- Hours (derived)
- 704 h
- Episodes (stated)
- 32,246 episodes
- Scale
- more than 720 hours of training demonstrations spanning diverse tabletop manipulation tasks
- Task
- Pick & place, Sorting & packing, Laundry & folding, Cleaning, Dishwashing, Electrical work +1 more
- Environment
- Indoor – lab
- Modality
- RGB, Proprioception, Actions, Language
- Capture
- Exocentric
- In frame
- Robot hands
- Captured in
- Undisclosed
- Embodiment
- bimanual YAM arms with grippers (robot_type bi_yam_follower)
- Frame rate
- 30 fps
- Resolution
- 640x360
- Formats
- LeRobot v3.0, Parquet, MP4
- Size
- about 2.35 TB on the Hub
- Hosted on
- Hugging Face
Figures are the publisher's own. Kinetic Blocks has not measured this dataset. Derived from meta/info.json: 76,046,658 frames at 30 fps. Frames are timesteps, not summed across the three cameras. The card says more than 720 hours across the full collection.
About this dataset
The bimanual training data behind Ai2's MolmoAct2 action reasoning models, merged into one LeRobot v3.0 repository. Two YAM arms work a tabletop through 34 tasks: spelling words and building shapes from blocks, lifting, rotating and flipping blocks, packing a container, plugging a charger into a phone and switching on a socket, folding clothes, clearing and stacking plates, scanning barcodes, scooping and weighing on a scale, sorting waste and loading a dishwasher rack, and untangling cables. Each episode has three AV1 camera streams (left, right and top) at 640x360 and 30 fps, joint and gripper state and action, and per-episode annotated instructions in meta/tasks_annotated.parquet.
Publisher's own task names: spell out words with blocks, form a tower, lift, rotate and flip blocks, pack container, plug charger to phone and switch on socket, clothes folding, clean dirty plates and stack, scan the item barcode, scoop and weigh on scale, sort waste and load the dishwasher rack, untangle cables.
What Kinetic Blocks did, and did not do
Indexed and linked from the publisher. Not hosted, verified or graded by Kinetic Blocks. The publisher's terms govern. The licence shown is the one stated on the publisher's page, linked above. Nothing on this page is a claim by Kinetic Blocks.
Want the verified ones too?
Request access to the marketplace, where verified listings are checked against their files and graded A+ to F, and your money waits until the data holds up.