CollisionCode/calvin_abc_d_lerobot_v2.1

Calvin ABC D (LeRobot v2.1 Format)

Dataset Description

This dataset is a converted version of clavin_abc_d.zip, now formatted for compatibility with the starVLA framework.

While the official Calvin dataset only provides zip files and HDF5 conversion methods, this version enables seamless training with modern imitation learning pipelines like Diffusion Policy, ACT, and VLA models within the LeRobot framework.

• Original Dataset: Calvin
• Conversion Tool: RoboTron-Mani, starVLA issue
• Format: LeRobot v0.3.3
• Robot: Panda Robot (7-DOF Arm + Gripper)

Dataset Structure

The dataset follows the standard LeRobot v2.1 structure:

• data/: Contains episode data in Parquet format.
• meta/: Contains task definitions, statistics, and camera configurations.
  • ⚠️ Note:: modality.json, stats_gr00t.json, steps_data_index.pkl are intermediate products generated during the StarVLA training pipeline
• videos/: (If included) High-quality task execution videos.

Key Observations/Actions

• observation.images.agentview: 200x200 agent-view camera RGB.
• observation.images.eye_in_hand: 84x84 wrist-mounted camera RGB.
• observation.state: Robot joint positions and gripper state.
• action: 7-DOF delta end-effector pose + gripper action.