Unitree Z1 teleoperation: a light arm for legged robots, with ROS 1, community ROS 2 and a LeRobot gap

The Unitree Z1 is a six-axis arm that weighs 4.3 kg (Air) or 4.5 kg (Pro), reaches 740 mm, carries 2 kg (Air) or 3 kg and more (Pro), and repeats to about 0.1 mm by Unitree's figure. It has position and force control, force feedback and collision detection, quick-change end actuators, and it is built to "cooperate with Aliengo or B1 Robot and other mobile robots". That combination, a light arm with a precise spec, is why it shows up on legged manipulators. The software around it is thinner than on a typical lab arm, so most of the work in teleoperating a Z1 is choosing, and owning, the control path.

The software stack, as it actually exists

How to teleoperate a Z1

Unitree does not document a dedicated leader arm or VR app for the Z1 in the material above, so every teleoperation setup is something you assemble. The practical options:

Putting a Z1 on a mobile base

Most Z1 teleoperation problems on mobile robots are integration problems that look like operator problems. Settle these before the first remote session:

Control loop and latency on a moving base

A teleoperation loop on a Z1 has two halves: the arm controller, reached over Ethernet and UDP, and whatever you put in front of it. Run that front end on a computer on the same robot, next to the arm, and keep the operator's link separate. If you record at 30 Hz, the rate of Unitree's public Z1 Dual dataset, each step is about 33 ms. On a legged or wheeled robot the operator's link usually goes over Wi-Fi, and Wi-Fi delay arrives in bursts. Clamp per-step joint motion on the robot side, filter the operator's targets, and give the operator a clear way to freeze the arm.

Our latency guide covers budgets by task type; start there before planning a remote Z1 program.

Data: Unitree's Z1 datasets and the LeRobot path

Unitree publishes Z1 data in LeRobot format. unitreerobotics/Z1_Dual_Dex1_StackBox_Dataset has 218 episodes and 117,266 frames at 30 fps, in LeRobot format v2.1, with one 640×480 high camera and a 14-value state and action: waist, shoulder, elbow, forearm roll, wrist angle, wrist rotate and gripper for each arm. In unitree_lerobot, teleoperation data is first stored as JSON per episode (color, depth, audio and a data file) and then converted to LeRobot format, with a small editor for trimming segments and deleting bad episodes. Its latest release moved to LeRobot dataset v3.0 and added π0.5 and GR00T policies.

Whichever path you take, write the stack into the dataset metadata. A Z1 driven through z1_ros on Noetic, through the community ROS 2 package, or straight through the SDK can produce data with different joint names, units, gripper conventions and timing. Unitree's own data uses names like kLeftWaist and kLeftWristRotate; a ros2_control setup will use whatever its URDF says. Map them once, in a script you keep with the data, rather than by hand at training time.

We deliver Z1 episodes as LeRobot, RLDS or HDF5, and as ROS 2 bags if the arm is part of a vehicle program. The data collection guide has the QA checklist.

What a good Z1 operator does

Z1 or a lab arm?

Choose the Z1 when weight and mounting decide: it is light enough for a legged or small mobile base, and Unitree positions it for its own quadrupeds and other mobile robots. For a fixed bench, an arm with a maintained LeRobot path is less work. The Trossen WidowX AI has an official LeRobot plugin, and Trossen's Mobile AI is a wheeled bimanual kit if mobility is the point. A UFACTORY xArm 6 brings maintained ROS 2 packages and more payload, on a heavier arm. The Z1 gives you a precise spec in a light package; you pay for it in integration.

How RoboPilots runs Z1 teleoperation

The robot stays with you. We agree the input path and control stack with you (ROS 1, community ROS 2 or SDK), install the operator client on your side of a VPN, and train a dedicated group of operators on your unit, your base and your task protocol before the first recorded episode.