Datasets:
G1 box into box (MuJoCo sim, scripted, 15 joints)
This dataset was created using LeRobot (format v3.0, lerobot 0.4.1).
The episodes are simulated. A MuJoCo Unitree G1 (29 DoF, gloved rubber hands, D405 wrist cameras) runs in a real2sim replica of the lab desk. It is driven over the same DDS + teleimager channels as the real robot. A code-as-policy script (g1lab.scripted.box_into_box) solves the task from the sim's privileged object poses through xr_teleoperate's IK. It squeezes the bubble-wrapped box between the palms, lifts it, turns the waist toward the cardboard box and lowers it inside. Only episodes that pass the sim's success check are kept. Every episode has a randomized start pose, speed, grasp parameters and object poses, and the sim adds visual randomization (--randomize).
- Task:
put the wrapped box inside the cardboard box - Episodes / frames: 100 / 48056 at 30 fps
- robot_type:
Unitree_G1_ArmsWaist_3Cam - Joints (15): the 14 arm joints (left, then right: shoulder pitch/roll/yaw, elbow, wrist roll/pitch/yaw) +
kWaistYaw(G1 motor 12)
Features
| key | type | shape | notes |
|---|---|---|---|
observation.state |
float32 | 15 | measured joint positions in sim (rad) |
action |
float32 | 15 | commanded joint targets (rad), as sent on rt/lowcmd |
observation.images.cam_head |
video (AV1, yuv420p) | 480x640x3 | rendered head camera |
observation.images.cam_left_wrist |
video | 480x640x3 | rendered left wrist camera |
observation.images.cam_right_wrist |
video | 480x640x3 | rendered right wrist camera |
The layout is the same as mobilab-udesc/g1_pick_cube_real_waist, so the two merge directly for real+sim training. The sim images are renders: expect a visual domain gap.
Good to know
- Appearance. The scene is a real2sim copy of the lab desk: cameras, object poses, colours, image blur and noise are calibrated against the 52 real episodes of
mobilab-udesc/g1_pick_cube_real. The wrapped box is textured from real camera frames. Every episode also has small random changes of lights, material tints and camera poses. - Motion. A scripted policy, not a human: reach, squeeze, lift, turn the waist, lower into the box, release, return. Speed, squeeze width, grasp height, lift clearance, waist angle and the start pose vary per episode. Episodes are 14-19 s long (mean 16 s) (the real teleoperated ones are about 11 s).
- Repeated frames. The recording PC rendered the three cameras a little below the 30 Hz recording rate, so about one frame in ten repeats the previous one.
- Waist. The waist yaw moves in every episode (range about 0.75 rad). The real set has waist = 0 throughout: train on a mix, not on the real set alone, if you want a 15-joint policy.
- Task text differs from the real set (
pick up cube.), which is xr_teleoperate's default text for the same task.
Citation
[More Information Needed]
- Downloads last month
- -