madang6/quickdraw-calvin-d-d
QuickDraw CALVIN D-to-D The official six-hour CALVIN environment-D play corpus, repackaged as LeRobot v2.1 for QuickDraw. Splits are made only at native rollout boundaries. train and validation are a deterministic 90/10 partition of upstream training rollouts; eval_offline preserves the official validation recordings. CALVIN's standard online D-to-D benchmark remains simulator evaluation. Each frame contains the unchanged 39-dimensional robot-plus-scene state, unchanged… See the full description on the dataset page: https://huggingface.co/datasets/madang6/quickdraw-calvin-d-d.
QuickDraw CALVIN D-to-D
The official six-hour CALVIN environment-D play corpus, repackaged as LeRobot v2.1 for QuickDraw. Splits are made only at native rollout boundaries. train and validation are a deterministic 90/10 partition of upstream training rollouts; eval_offline preserves the official validation recordings. CALVIN's standard online D-to-D benchmark remains simulator evaluation.
Each frame contains the unchanged 39-dimensional robot-plus-scene state, unchanged 7-dimensional relative action, the absolute action as an auxiliary field, and synchronized static and gripper RGB. Language spans and canonical CALVIN task names are retained as annotations; unlabeled play remains available rather than being discarded.
Derived from CALVIN, released under the MIT license.