red_block_only

Bimanual manipulation demonstrations recorded with Raiden and exported to the LeRobot v3.0 format.

Dataset

• Codebase version: v3.0
• Robot type: yam
• Episodes: 47
• Frames: 20822
• Tasks: 1
• FPS: 30

Features

Key	dtype	Shape
action	float32	(7,)
action.ee_pose	float32	(13,)
episode_index	int64	(1,)
frame_index	int64	(1,)
index	int64	(1,)
observation.effort	float32	(7,)
observation.extrinsics.left_wrist_camera	float32	(16,)
observation.extrinsics.scene_camera	float32	(16,)
observation.images.depth_scene_camera	video	(480, 640, 1)
observation.images.left_wrist_camera	video	(480, 640, 3)
observation.images.scene_camera	video	(480, 640, 3)
observation.intrinsics.left_wrist_camera	float32	(9,)
observation.intrinsics.scene_camera	float32	(9,)
observation.leader.joint_effort	float32	(6,)
observation.leader.joint_position	float32	(6,)
observation.leader.joint_velocity	float32	(6,)
observation.state	float32	(7,)
observation.state.ee_pose	float32	(13,)
observation.timestamp_ns	int64	(1,)
observation.velocity	float32	(7,)
task_index	int64	(1,)
timestamp	float32	(1,)

Camera intrinsics are stored per frame as observation.intrinsics.<camera> (flattened 3x3), and wrist-camera extrinsics as observation.extrinsics.<camera>. Depth is 12-bit logarithmically quantized in the video stream; a lossless first frame of each episode (RGB PNG + uint16 depth PNG + calibration JSON) is archived under meta/keyframes/ for pose estimation.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("YzyLmc/red_block_only")
print(dataset[0].keys())