What Unitree humanoids can do beyond the viral videos: G1 research and H2 specs

Unitree's G1 is one of the most widely used humanoids in university research, and the papers show it can jump, kick, climb stairs and pick up objects. The same papers also show the limits: robots overheating and breaking during agile moves, pick-up success rates around 75–80%, and stair walking that fails outside the lab. For the larger H2, there is so far little published evidence beyond Unitree's own specifications.

Why the viral videos are hard to read

Most Unitree clips online show choreography: a dance, a kung fu routine or a flip that has been recorded, tuned and rehearsed for one specific setting. That work is real, but it says little about what the robot does when something unexpected happens. A routine that works perfectly on a familiar floor may never have been tried on a carpeted stage or under bright lights.

Research papers are a better guide because they describe the method, the hardware and, in the better ones, what went wrong. Several groups have published work using the G1 as their test platform. None of them is an independent product test, and each one runs the authors' own research software rather than what ships with the robot. Even so, they are the most detailed public record of what the hardware can and cannot do.

What Unitree itself says about the G1

On its G1 product page, Unitree lists a standing size of 1,320 × 450 × 200 mm, a weight of about 35 kg with battery, and a battery life of about two hours. The standard G1 has 23 degrees of freedom (independently moving joints). The G1 EDU version can be configured with 23 to 43, including an optional three-fingered Dex3-1 hand. Unitree lists a maximum knee torque of 90 N·m for the G1 and 120 N·m for the EDU. Arm payload is about 2 kg and 3 kg respectively.

Unitree lists a depth camera, a 3D LiDAR, a four-microphone array and a 5 W speaker. On the same page it gives a price of US$13,500 for the standard G1 and says EDU pricing is available from sales. Two details matter for anyone hiring one:

• Secondary development. Unitree marks the standard G1 as not supporting secondary development and the EDU as supporting it. Custom programming generally depends on which version a supplier owns.
• The manufacturer's own caveats. The page says some of the sample functions shown are still being developed and tested. It describes the system as extremely powerful and asks users to keep a safe distance between robot and people. It also says the humanoid industry is at an early stage of exploration.

For how these figures appear in our catalogue, see the Unitree G1 robot page.

Agile moves: the ASAP paper

In February 2025, researchers from Carnegie Mellon University and NVIDIA published ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. The team trained motion-tracking policies in simulation from human motion data. They then used data from the real G1 to correct for differences between the simulated and real physics.

On the real robot, the authors show kicks, a forward jump of about 1.5 metres, single-leg balance and single-leg jumps. They also show recreations of athletes' celebrations, including a jump with a 180-degree turn. They report lower tracking error than the comparison methods.

The paper is just as useful for what it admits. The authors write that two G1 robots broke during data collection because their motors overheated during the dynamic motions. As a result they reduced the scope of their correction model from the full body to the ankles. For an event, the lesson is simple: highly agile moves put real strain on the hardware. A supplier who limits how often a robot performs its most dramatic moves is being sensible, not cautious for no reason.

Teleoperated work: the HOMIE paper

Also in February 2025 (revised in April), the Shanghai AI Laboratory and The Chinese University of Hong Kong published HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit. Here a human operator drives the robot. A pedal controls walking, an exoskeleton arm mirrors the operator's arm movements, and sensor gloves control the hands. A learned policy keeps the legs balanced. The main platform was the G1, with a Fourier GR-1 used to show the method also works on another robot.

The authors report that their system completed tasks in about half the time of earlier teleoperation setups, with cockpit hardware costing around US$500. When policies trained on the operators' demonstrations ran autonomously, the full paper reports 73.3% success on a squat-and-pick task and 80% on pick-and-place, each over 15 trials. The authors also state that the robot struggled to walk stably on real stairs even though it succeeded in simulation. The system has no force feedback for the operator.

This matters because much of what audiences see as a robot "handing out" items or "interacting" is teleoperated. Our guide on teleoperated versus autonomous robots explains what that means for your audience.

Stairs and chairs: the VideoMimic paper

In May 2025 (revised in August), a UC Berkeley team published Visual Imitation Enables Contextual Humanoid Control, known as VideoMimic. The pipeline extracts human movements and the surrounding scene from ordinary videos, rebuilds them in simulation and trains a single control policy. On a 23-degree-of-freedom G1 running the policy onboard, the robot climbed and descended stairs and sat down on and stood up from chairs and benches.

The full paper is candid about the limits:

• The LiDAR height map the robot uses is good enough for terrain and chairs, but not for precise contact, manipulation or overhanging obstacles.
• The policy sometimes falls back on recovery behaviours, which makes its movements jerky.
• Training used only 123 video clips.
• The paper gives no numerical success rates for the real-world tests.

What we could not find for the H2

On its H2 product page, Unitree lists a height of 1,820 mm, a weight of about 70 kg with battery and 31 degrees of freedom. It gives leg joint torque of up to 360 N·m, an arm payload of about 7 kg rated and 15 kg peak, and a runtime of about three hours. The listed price is US$29,900 for the standard H2. The same page carries the same safe-distance warning as the G1.

While researching this article we did not find independent papers or tests of the H2 comparable to the G1 work above. Until such work appears, its capabilities rest on Unitree's specifications and its own videos. A robot twice the weight of a G1 also raises the stakes if it falls or makes contact with someone. Our earlier article on H2, H2-X and R1 specs and prices covers the published figures in more detail.

Separating choreography from capability

Taken together, the sources suggest three tiers of what a Unitree humanoid can do:

1. Rehearsed routines. Dances, kicks, jumps and martial-arts sequences that have been recorded and tuned in advance. These can look spectacular and can be repeated, but they strain the motors and assume a known floor and space.
2. Operator-driven tasks. Walking among people, gesturing and handing over light objects, usually with a skilled operator in control. The arm payloads Unitree lists, about 2–3 kg for the G1, set firm limits.
3. Autonomous skills. Picking up objects, climbing stairs and sitting on chairs have been shown in research. But they rely on research software, work less than perfectly in the authors' own trials, and are not features you can assume a rental unit has.

What this means for event organisers

• Ask which tier you are buying. In your brief, say whether you want a rehearsed routine, an operated presence or something autonomous, and ask the supplier which one they are actually offering. Our guide on what to put in a request to a robot supplier has a checklist.
• Budget time for choreography. Any new routine needs development and testing on hardware, and a standard G1 may not support custom programming at all. See custom moves and choreography lead times.
• Expect dramatic moves to be rationed. Overheating is documented in research, so a short, repeated show slot is more realistic than continuous acrobatics.
• Plan for distance, not contact. Unitree itself asks for a safe distance from people. Build that into the stage layout using our guide to safety at public events with robots.
• Treat H2 claims as unverified. Until independent results are published, judge the H2 on what a supplier can show you live, not on launch videos.

Frequently asked questions

Can a hired Unitree G1 do the flips and kicks seen in research videos?

Not automatically. The research papers ran their own software on the robot, and the ASAP authors report that two G1 robots broke from motor overheating during agile motion trials. Whether a rental unit can perform a specific move depends on the supplier's software, the version of the robot and how much testing they have done. Ask the supplier to show the exact routine on their own hardware before you build a show around it.

What is the difference between the Unitree G1 and the G1 EDU?

Unitree lists the standard G1 with 23 degrees of freedom and no support for secondary development. The G1 EDU can be configured with 23 to 43 degrees of freedom, including optional dexterous hands, and supports secondary development. It has higher listed knee torque and arm payload and a longer warranty. For events this matters because custom programming generally needs the EDU version.

Is the Unitree G1 autonomous at events?

Usually only partly. Balance and walking are handled by the robot's controller, but tasks such as walking a route, gesturing at guests or handing over objects are commonly operated by a person or played back as a prepared routine. Research shows the G1 can do some tasks autonomously, such as picking up objects, but with success rates of roughly 73–80% in one study's trials, which is not reliable enough to promise to an audience.

Has the Unitree H2 been independently tested?

When this article was researched, we did not find independent research papers or tests of the H2 comparable to the published work on the G1. Its capabilities are therefore described by Unitree's product page, which lists a height of 1,820 mm, about 70 kg, 31 degrees of freedom and around three hours of runtime. Treat claims about the H2 as manufacturer statements until independent results appear.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-11-03. Spotted an error? Write to info@humanoid4rent.com.