Traly/calvin_abc_d-lerobot

CALVIN ABC-D LeRobot v3.0

This dataset is a LeRobot v3.0 conversion of the CALVIN ABC-D split.

Key properties:

• Format: LeRobot v3.0
• FPS: 30
• Robot: Franka Panda
• Episodes: 17,870
• Frames: 1,071,743
• Tasks: 389
• Images are stored as H.264 videos.
• Both absolute end-effector actions and relative end-effector displacement actions are included.

Action Convention

The default action key is the absolute end-effector action and is identical to action.absolute.

All action vectors are 7D:

[x, y, z, rx, ry, rz, gripper_action]

where rx, ry, rz are Euler orientation components.

Feature Keys

Key	Shape	Type	Description
observation.images.top	[200, 200, 3]	video, H.264, 30 FPS	Static/top RGB camera. LeRobot returns this as CHW float tensor when loaded.
observation.images.wrist	[84, 84, 3]	video, H.264, 30 FPS	Wrist RGB camera. LeRobot returns this as CHW float tensor when loaded.
observation.state	[15]	float32	Robot state: EE position, EE Euler orientation, gripper width, 7 joint positions, and gripper action.
observation.environment_state	[24]	float32	CALVIN scene state: door, drawer, button, switch, lights, and block poses.
action	[7]	float32	Default training action. This is the absolute EE action and equals action.absolute.
action.absolute	[7]	float32	Absolute EE action: [ee_position_x, ee_position_y, ee_position_z, ee_orientation_rx, ee_orientation_ry, ee_orientation_rz, gripper_action].
action.relative	[7]	float32	Relative EE displacement action: [delta_ee_position_x, delta_ee_position_y, delta_ee_position_z, delta_ee_orientation_rx, delta_ee_orientation_ry, delta_ee_orientation_rz, gripper_action].
timestamp	[1]	float32	Frame timestamp in seconds.
frame_index	[1]	int64	Frame index inside the episode.
episode_index	[1]	int64	Episode index.
index	[1]	int64	Global frame index.
task_index	[1]	int64	Task index. Task strings are stored in meta/tasks.parquet and exposed as task by LeRobotDataset.

observation.state Axes

[
  ee_position_x,
  ee_position_y,
  ee_position_z,
  ee_orientation_rx,
  ee_orientation_ry,
  ee_orientation_rz,
  gripper_width,
  joint_position_00,
  joint_position_01,
  joint_position_02,
  joint_position_03,
  joint_position_04,
  joint_position_05,
  joint_position_06,
  gripper_action
]

observation.environment_state Axes

[
  sliding_door,
  drawer,
  button,
  switch,
  lightbulb,
  green_light,
  red_block_x,
  red_block_y,
  red_block_z,
  red_block_rx,
  red_block_ry,
  red_block_rz,
  blue_block_x,
  blue_block_y,
  blue_block_z,
  blue_block_rx,
  blue_block_ry,
  blue_block_rz,
  pink_block_x,
  pink_block_y,
  pink_block_z,
  pink_block_rx,
  pink_block_ry,
  pink_block_rz
]

Loading Example

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="Traly/calvin_abc_d-lerobot",
    video_backend="pyav",
)

sample = dataset[0]
absolute_action = sample["action.absolute"]
relative_action = sample["action.relative"]