Pre-launch Submit your use case
qtvue
← All platforms

Unitree G1-D

Contact for pricing

The dual-arm teleoperation rig for embodied-AI teams collecting training datasets — sub-100 ms latency.

Data/training platform·for Embodied-AI teams

Specs

What it actually does.

Spec sheet

Latency
< 100 ms (operator -> robot)
Arms
Dual dexterous arms (G1 EDU spec)
Base
G1 humanoid base
Recording
Synchronized RGB-D + joint state + timestamps
Output format
ROS 2 bags, LeRobot-compatible datasets
Compute
Onboard + workstation-grade offboard option

Use cases

  • Imitation learning dataset collection
  • Teleoperation research (shared autonomy, RLHF)
  • Embodied-AI training data at scale
  • VR/leader-follower manipulation studies

Honest callouts

  • · Sub-100 ms latency is the working number for natural-feeling teleoperation. Marketing claims of <50 ms are under specific network conditions.
  • · Output is ROS 2 bags and LeRobot-compatible datasets — plug into any modern VLA training pipeline.
  • · Indoor only. The G1-D is built for the lab, not the field.
g1d_record.py
python
# G1-D — record a teleop episode
from unitree_sdk2py.core.channel import ChannelFactoryInitialize
from unitree_sdk2py.g1_d.teleop import TeleopRecorder

ChannelFactoryInitialize(0, "eth0")
rec = TeleopRecorder(out_dir="./episodes/2026-06-22/")

rec.start_episode(task="pick_cube")
while operator_active:
    rec.record_step({
      "rgb_left":  cam_left.frame(),
      "rgb_right": cam_right.frame(),
      "depth":     cam_depth.frame(),
      "joints":    arm.state(),
      "gripper":   gripper.state(),
    })
rec.stop_episode()  # writes a LeRobot-compatible dataset

Get one

Submit your use case.

Tell us what you want to do and we'll tell you whether this platform is the right one (and if not, which is).

Submit your use case →

Other platforms

Compare the range.