Episodes Preview imago Visualizer
1k episodes · 10 fps · 1 camera · 128×128 av1

LamportRobotics/imago_pack_can_1k

1000 scripted demonstrations of one task in simulation, LeRobot v3 layout, 10 Hz (one frame per picture).

Task: Pick the can off the counter and put it in the tote.

The can stands somewhere new each time (x 0.33..0.44 m, y -0.32..-0.14 m in front of the robot) and weighs 0.25..0.6 kg. A scripted expert solves the grasp for that position and carries the can over the tote. 1457 attempts were run; 430 were out of reach for the expert and 8 missed the tote. Those are not in the set. An episode is kept only if the can ends in the tote and the robot is still upright.

Features

  • observation.state: 14 joint positions, rad
  • action: 14 joint position targets, rad (what the expert commanded)
  • observation.images.eye: the camera in the visor, 128 x 128 RGB at 10 Hz. The camera turns with the waist.

Joints, in order: waist_yaw, right_shoulder_pitch, right_shoulder_roll, right_shoulder_yaw, right_elbow, right_wrist_roll, right_wrist_flex, right_wrist_dev, right_index, right_middle, right_ring, right_little, right_thumb_swing, right_thumb.

The robot

Imago is Lamport's humanoid, a 1.50 m, 48 kg simulation model with five-finger hands, built in MuJoCo. It exists only in simulation. The task uses 14 of its 42 driven joints: waist yaw, the right arm (7) and the right hand (6). Joint zero is the standing pose.

Numbers

episodes 1000
frames 136000
rate 10 Hz
success rate of the expert 1000/1457 attempts kept

Load with LeRobotDataset("LamportRobotics/imago_pack_can_1k").

Generated by Lamport Robotics. The policy trained on these demonstrations is LamportRobotics/imago-pack-can.

Downloads last month
-

Models trained or fine-tuned on LamportRobotics/imago_pack_can_1k