kisarakira/robotwin2-lingbot-vla-v2-lerobot-v3

RoboTwin2-LingBot-VLA-v2-LeRobot-v3

This dataset is a LeRobot v3.0 formatted version of RoboTwin 2.0, prepared for post-training of LingBot-VLA-v2.

Overview

Following the RoboTwin post-training configuration provided by LingBot-VLA-v2, this dataset converts the original RoboTwin 2.0 HDF5 data into the LeRobot v3.0 format.

The converted dataset can be directly used for RoboTwin post-training with LingBot-VLA-v2.

Conversion Details

The conversion pipeline includes:

• —Converting RoboTwin 2.0 HDF5 trajectories into LeRobot v3.0 format.
• —Organizing episodes and metadata following the LeRobot dataset structure.
• —Renaming task identifiers to match the task names defined in:

assets/training_data/robotwin.txt

from the LingBot-VLA-v2 repository.

Usage with LingBot-VLA-v2

This dataset is designed to be compatible with the RoboTwin post-training pipeline of LingBot-VLA-v2.

Reference:

https://github.com/Robbyant/lingbot-vla-v2

The task names are aligned with:

assets/training_data/robotwin.txt

Configuration Note

The default camera configuration in LingBot-VLA-v2:

configs/robot_configs/robotwin.yaml

may need to be modified to match the original camera configuration of the RoboTwin 2.0 dataset.

Please ensure that the camera names and image feature definitions are consistent before training.

Dataset Format

This dataset follows the LeRobot v3.0 format.

Directory structure:

data/
meta/
videos/

Source Dataset

This dataset is derived from:

• —RoboTwin 2.0
• —LingBot-VLA-v2 training configuration

Please follow the original licenses of the corresponding projects.

Acknowledgement

Thanks to the RoboTwin and LingBot-VLA-v2 authors for their open-source contributions.