---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- LIBERO
- MolmoAct2
configs:
- config_name: default
  data_files: data/*/*.parquet
---

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

## MolmoAct2-LIBERO Dataset

This repository contains the merged LeRobot LIBERO dataset used for MolmoAct2 finetuning experiments. It combines all four LIBERO suites used in our finetuning setup: LIBERO-Spatial, LIBERO-Goal, LIBERO-Object, and LIBERO-Long / LIBERO-10.

The merged dataset contains 1,693 episodes and 273,465 frame-level training examples at 10 FPS. It is intended for finetuning only. We did not use this dataset in MolmoAct2 pretraining or post-training.

## Language Annotations

No additional language annotations are added in this release. The dataset includes only the standard LeRobot task metadata in [`meta/tasks.parquet`](meta/tasks.parquet), referenced by each frame through `task_index`. There is no `meta/tasks_annotated.parquet` file.

## Normalization Statistics

For MolmoAct2 training, 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 normalization scheme worked well for MolmoAct2 finetuning.

## Dataset Structure

The dataset follows the standard LeRobot v3 layout:

- [`meta/info.json`](meta/info.json): dataset metadata and feature schema
- [`meta/stats.json`](meta/stats.json): normalization statistics
- [`meta/tasks.parquet`](meta/tasks.parquet): standard LeRobot task table
- `data/chunk-*/file-*.parquet`: frame-level robot observations and actions

## Citation

```bibtex
@misc{fang2026molmoact2actionreasoningmodels,
      title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
      author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
      year={2026},
      eprint={2605.02881},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.02881},
}
```