Unitree H2 teleoperation: human-height reach, the EDU requirement and the H2 Plus GR00T stack

Unveiled in October 2025, the Unitree H2 is a full-size humanoid with a list price under $30,000. It stands 1.82 m, weighs about 70 kg, and has 31 joints including a 3-DoF waist and a 2-DoF head. Its arms are rated for about 7 kg each. That combination matters for data: a robot at adult height reaches the same shelves, counters and workbenches a person does, so the demonstrations you record are for the environment you actually want to automate. The teleoperation story is younger than the G1's, though, and the gaps are worth knowing before you plan a data program around it.

Three H2s, and only two you can develop on

Unitree sells the H2 in three forms, and the spec table settles which ones you can teleoperate through your own code. Secondary development is marked "YES" for the H2 EDU and for the H2 Plus, and left blank for the standard H2.

As of October 2026, foreign-produced advanced robotic devices are on the FCC Covered List under FCC public notice DA 26-786 (July 28, 2026), unless the Department of War grants them a Conditional Approval. The FCC definition covers mobile robots, including humanoids, quadrupeds and autonomous mobile robots, that weigh over 4.4 lb with any base or dock and have a sensor, network connectivity and control software, when they do not qualify as US-made under the federal Buy American test, whatever the nationality of the maker. Covered models cannot receive new FCC equipment authorizations, which most devices need before they can be imported or sold in the US, while the FCC says models it authorized earlier can still be imported, sold and used; US buyers should confirm the current status with the seller.

How to drive an H2 remotely today

The H2 EDU uses the same unitree_sdk2 and DDS plumbing as the rest of Unitree's humanoids, so most of what works on a G1 carries over in principle. In practice, two stacks are usable right now.

One gap to plan around: an open issue on xr_teleoperate, filed on July 26, 2026, asks when H2 head data can be collected in sync with the rest of the stream. It had no maintainer reply when we checked in October 2026. If your policy needs the head joints in the observation, confirm with Unitree or record them through your own SDK node before you start.

What the extra size changes for the operator

Height is the reason to buy an H2, and it changes the operator's job in three ways.

Whole-body versus upper-body data on the H2

Most H2 teleoperation today is upper-body teleop with commanded walking: the arms and hands follow the operator, and the legs follow the joysticks through the robot's own locomotion controller. That is the right default for counter and shelf work. The waist sits in between. If the operator leans the torso through the stack, those three joints belong in the action vector; if the waist stays fixed, leave them out and say so in the dataset card.

True whole-body data, where the legs are part of the demonstration (squatting to lift, stepping while carrying), needs a whole-body controller and a capture device that tracks the operator's body, not just the hands. On the H2 that work is early. Unitree's unitree_rl_gym lists the Go2, H1, H1_2 and G1 but not the H2, while the C++ simulator in unitree_mujoco does include an H2 configuration and uses the same DDS interface as the real robot (domain 1 on loopback for simulation, domain 0 on the real network). That makes MuJoCo the place to rehearse a teleop protocol before an operator touches the real robot.

Hands: Dex3, five-finger options or SharpaWave

On an H2 EDU you choose from Unitree's dexterous-hand options; on the H2 Plus you can fit SharpaWave hands with 22 active DoF, more than 1,000 tactile pixels per fingertip, 150 N grip force and ±1 mm fingertip repeatability. More fingers do not automatically mean better data. Hand tracking from a headset maps the operator's hand onto the robot's, and the more joints the robot hand has, the more the operator's technique and the retargeting decide the result. For pick, place and carry tasks, a simpler hand often yields a higher clean-episode rate. For tool use and in-hand work, the tactile hand is the reason to buy the Plus, and the tactile streams should be recorded from day one even if the first policy ignores them.

Latency for a full-size robot

The control problem is the same as on any Unitree humanoid: a fast local loop (locomotion, IK, safety limits) next to the robot, and a slow loop to the operator over the network. xr_teleoperate is built for a headset and host on the same network, so a remote operator connects to a host on your site rather than to the robot directly. What changes on the H2 is the cost of an overshoot. A heavy arm moving a 7 kg load a few hundred milliseconds behind the operator's hand is harder to stop cleanly than a small arm with a cup. Keep speed limits tighter on a full-size robot once round-trip delay passes about 100 ms, and move fine work from direct control to supervision above about 200 ms. Our latency guide covers the budgets by task.

What a good H2 operator does

H2 versus H1 and G1

Within Unitree's range, the H2 sits between two older robots. The Unitree G1 is smaller and cheaper, with the deepest open ecosystem and hundreds of public datasets; it is the better choice when you want many robots or tabletop tasks. The Unitree H1 is the full-size robot behind many research teleop papers, with more published research code than the H2 but a higher list price. The H2 is the choice when you need adult reach and real payload in a robot you can buy several of, and you accept that some tooling is still catching up.

Takeover and fallback

If you deploy an H2 running a policy, plan the takeover path before the first shift: who watches the robot, what triggers a handover, and which interface the operator uses to finish the task. Using the same stack for takeover as for data collection means each correction is recorded in the same format and can go straight into the next training run. Our human-in-the-loop guide covers the escalation design for humanoids.

How RoboPilots runs H2 teleoperation

The robot stays at your site. We agree the stack with you (xr_teleoperate, Isaac TeleOp on an H2 Plus, or your own), set up the operator link through a host on your network, and train a dedicated group of operators on your unit, your hands and your task protocol, including the fall and stop procedures for a full-size robot. Several humanoid makers have sent us robots to test our operating protocols; those protocols are the starting point, and your H2 program gets its own.