UFACTORY xArm 6 teleoperation: leader arms, servo control and clean training data

The UFACTORY xArm 6 sits in a sweet spot for robot learning: a 700 mm reach and 5 kg payload, ±0.1 mm repeatability, an open Python SDK and maintained ROS 2 packages, at a price a lab can buy several of. That is why it shows up in so many teleoperation setups. It is also why teams run into the same three problems with it: the input device, the control loop, and getting the data into a training pipeline.

Four ways to drive an xArm 6 remotely

Every option below ends at the same place: a stream of joint or Cartesian targets sent to the arm through the xArm Python SDK or the xarm_ros2 driver. What changes is how natural the operator's motion is, and how learnable the resulting trajectories are.

The control loop: servo modes, rates and latency

For teleoperation you want streaming control, not point-to-point moves. The SDK provides three families: servo streaming of joint angles (set_servo_angle_j) or Cartesian poses (set_servo_cartesian), joint and Cartesian velocity control (vc_set_joint_velocity, vc_set_cartesian_velocity), and online trajectory planning that re-plans smoothly when a new target arrives. Leader arms usually map to joint servo; 3D mice and VR map more naturally to Cartesian velocity. One catch: UFACTORY's servo mode guide recommends streaming at 50 to 200 Hz (250 Hz at most) and warns that below 50 Hz motion may not be continuous.

Most xArm learning setups record at 10 to 30 Hz, because that is what policies such as ACT, Diffusion Policy and smolVLA are typically trained at. That mismatch matters. If your policy or your recording loop runs at 20 Hz, either interpolate its targets up to 50 Hz or more on the robot side before calling servo mode, or switch to the online trajectory planning modes (6 for joints, 7 for Cartesian), where the controller plans between targets itself and the rate requirement is lower. At 20 Hz each control step is 50 ms, and that number also sets the latency budget for a remote operator:

Two things narrow the gap. First, run the operator's client close to the robot over a VPN instead of streaming a remote desktop: remote desktop re-encodes and buffers the whole screen, which adds delay you cannot tune away. Second, keep a local loop on the robot side that clamps velocity and smooths targets, so a delayed command never turns into a jerk. More on budgets by task in our latency guide.

Getting xArm data into LeRobot

LeRobot itself does not include an xArm driver: its built-in robots are the SO-series, Koch, OpenArm, Reachy 2, Unitree G1 and a few others. UFACTORY fills the gap with its own plugin, lerobot_robot_ufactory, which connects xArm 6 and xArm 7 to LeRobot and ships teleop modes for GELLO, a SpaceMouse, the Pika handheld tracker and UMI. It pins a specific LeRobot release, so check it matches the version your training code uses. Teams with a custom setup (extra cameras, a depth sensor, their own gripper) often write their own Robot class around the xArm SDK instead; it is a few hundred lines, and after that LeRobot recording, visualization and training work as usual.

There is public xArm data to compare against: the Hugging Face Hub has hundreds of datasets that mention the xArm, including lerobot/utokyo_xarm_pick_and_place, a real-robot set ported from Open X-Embodiment. Our data collection guide covers formats and quality checks in more depth.

What a good xArm operator does differently

xArm 6, xArm 7 or Lite 6?

The xArm 7 adds a seventh joint, which gives the arm an elbow it can move without moving the gripper. It is useful for reaching around obstacles and for human-like bimanual setups, but it makes leader-arm mapping and inverse kinematics slightly more involved. The Lite 6 (600 g payload, 440 mm reach, ±0.5 mm repeatability per UFACTORY) is a cheaper desktop arm, good for teaching and coarse tasks but not for sub-millimetre work. For most single-arm manipulation datasets, the xArm 6 is the straightforward choice.

How RoboPilots runs xArm teleoperation

The arm stays in your lab. We agree the input device with you (your leader arm or 3D mouse, or one we set up), install the operator client on your side of a VPN, and train a dedicated group of operators on your unit and your task protocol before they record a single episode.