MolmoAct2-BimanualYAM-Dataset

This dataset was created using LeRobot. MolmoAct2-BimanualYAM Dataset This repository is the merged ckpt / merged LeRobot dataset artifact for the MolmoAct2-BimanualYAM Dataset, a large-scale collection of bimanual robot manipulation demonstrations collected for MolmoAct2. Across the full collection, MolmoAct2-BimanualYAM contains more than 720 hours of training demonstrations spanning diverse tabletop manipulation tasks. Dataset Description **Homepage:** [More Information Needed] **Paper:** [More Information Needed] **License:** apache-2.0 Language Annotations This dataset includes annotated language instructions in `meta/tasks_annotated.parquet`. The file is indexed by `episode_index` and has a `task` column containing our per-episode annotated instruction. The standard LeRobot loader resolves a frame's language instruction through `task_index`: each data row stores a `task_index`, which is looked up in `meta/tasks.parquet`. When you use these annotations, load `meta/tasks_annotated.parquet` and look up the current `episode_index` instead. If no valid annotated row is available, fall back to the standard LeRobot task. Normalization Statistics For MolmoAct2 training, we did not train directly from this single merged LeRobot repository. The training pipeline used a custom dataloader that takes multiple LeRobot datasets as separate sources and composes them at load time. Because the training mixture can be defined from arbitrary combinations of source datasets, quantile-based normalization statistics are estimated by taking weighted averages of per-source dataset quantiles, instead of recomputing exact global quantiles for every possible dataset mixture. This makes it practical to reuse source-level statistics across different combinations of datasets. Because quantiles are nonlinear, weighted averages of per-source quantiles are not identical to exact quantiles computed from the fully merged dataset, so this approximation can introduce some bias. In practice, this n