DiscoDemo-Stage2_GenData-StackCube

Demonstrations for the StackCube task (stack the red cube on the blue cube), generated in simulation by the DiscoDemo generator (diversity weight α = 0.3) of DiscoDemo. Only successful, safety-filtered episodes are kept.

Stage	Repository
Generator (RL policy)	DAVIAN-Robotics/DiscoDemo-Stage1_RL-StackCube
Generated dataset	DAVIAN-Robotics/DiscoDemo-Stage2_GenData-StackCube (this repository)
Imitation policy (π0.5)	DAVIAN-Robotics/DiscoDemo-Stage3_SFT-StackCube

Contents

Format	LeRobot v3.0
Episodes	3,000
Frames	510,837
Rate	20 Hz
Cameras	over_shoulder_left_camera, wrist_cam (480 × 640 RGB)
Language instruction	"stack the red cube on the blue cube"

• action (8): absolute joint-position targets for the 7 arm joints and a binary gripper command (0 open, 1 close).
• observation.state (8): the 7 arm joint positions and the gripper state.
• observation.velocity (8): the corresponding velocities.
• observation.skill_z (3): the skill vector the generator was conditioned on for the episode.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("DAVIAN-Robotics/DiscoDemo-Stage2_GenData-StackCube", revision="v3.0")

Citation

@article{park2026discodemo,
  title   = {DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning},
  author  = {Park, Minho and Kim, Kinam and Kim, Donghu and Lee, Byungkun and Hwang, Dongyoon and Shin, Yongjae and Hyung, Junha and Lee, Hojoon and Choo, Jaegul},
  journal = {arXiv preprint},
  year    = {2026}
}