mj_pick_block — Expert Pick-Block Demonstrations

300 episodes of a Franka Panda arm picking a cube in a standalone MuJoCo scene, recorded with LeRobot. Demonstrations are generated by a hardcoded expert policy, so the dataset is fully reproducible and label-clean for imitation learning.

Task

pick up the block — hover above the cube, descend to it, close the gripper, and lift it clear of the table.

Collection setup

• Simulator: MuJoCo
• Robot: Franka Emika Panda, single arm (robot_type: panda)
• Control: translation-only Cartesian control at 50 Hz, fixed top-down gripper orientation
• Expert: a privileged scripted policy that reads ground-truth cube and
  end-effector positions and runs a 4-phase state machine:
  1. Approach — move to 5 cm directly above the cube center
  2. Descend — move down to the cube center
  3. Close — close the gripper and hold for 1 s
  4. Lift — raise the end-effector to 50 cm world height, then end

Features

Key	Dtype	Shape	Description
observation.image.front	image	3×224×224	front scene camera (RGB)
observation.image.wrist	image	3×224×224	wrist camera (RGB)
observation.state	float32	(8,)	7 arm joint positions + 1 gripper finger position (m)
action	float32	(4,)	[dx, dy, dz, gripper]

Action space [dx, dy, dz, gripper]:

• dx, dy, dz — desired world-frame end-effector position delta for the step, in meters (clamped to ±2 cm/step), resolved to joint torques by a Jacobian-based Cartesian controller.
• gripper — target finger opening, 0.0 = fully closed, 1.0 = fully open.

Dataset structure

• Episodes: 300
• Frames per episode: variable (episode ends when the expert finishes the lift phase)
• Control frequency: 50 Hz