Unitree H1 Robot

How much does the Unitree H1 cost in 2026?

The listed figure is $90,000 on Unitree's own store, checked 30 August 2026, under a product title that literally reads "contact us for the real price". There is no online checkout. You email sales and get a quote.

Resellers are where the real numbers live, and they are higher. The largest US dealer lists the H1 at $99,900 and shows it sold out with a notify-me button. Another US seller lists the H1-2 variant at $122,000 with an active cart. In Europe, a French distributor quotes €93,496 excluding VAT, or €112,195 including it, with delivery times on request. A US listing at $94,905 appears in search results but we could not verify it on the page itself.

What that spread tells you is that the H1 is being sold out of remaining inventory at whatever the channel will bear, not manufactured to order at a list price. The $150,000 figure that appears on some aggregator sites has no primary source and looks like an error; the 2023 launch coverage put the robot below $90,000.

The comparison that actually matters is internal. The H2, launched October 2025, is $29,900 on the same store — 180 cm, 31 degrees of freedom, 360 N·m joint torque, 2,070 TOPS of compute, a bionic face and over-the-air updates. The H2 Plus is $100,000. The G1 is $13,500. Against that ladder, a $90,000-to-$122,000 H1 with 18 to 19 degrees of freedom is priced as a legacy artifact rather than as a product.

What does the H1 price include, and what does it exclude?

Included, per the dealer listings: the robot, a remote control and one battery. Distributor documentation adds the 3D LiDAR and depth camera, ROS compatibility, remote, app and PC control, Unitree's APIs and SDKs, and pre-shipment quality control.

Excluded, and this is where the budget goes:

Hands. The base robot has none. Unitree's Dex3-1 three-finger force-control hand is around $6,500 from a US dealer, an H1-2 dexterous hand is listed at $10,000, and Inspire-class hands run to roughly $12,000. Across the Unitree ecosystem, hand options span about $6,500 to $19,000 per configuration.

Compute. The standard platform is an Intel Core i5; an i7 or an NVIDIA Jetson Orin NX are upgrade options. Neither upgrade is priced by any seller we found.

Batteries. One pack ships with the robot. Spares list around $1,500 at resellers, and you will want two or three for a working lab day. One caution: reseller listings describe the pack as 28.8 V and 15,000 mAh, which computes to 432 Wh, while Unitree's own specification says 15 Ah at up to 67.2 V, or 864 Wh. Those cannot both be right. Confirm the pack chemistry and capacity before ordering spares.

Warranty. Not stated on any H1 listing we could find, from Unitree or from any dealer. That is unusual and worth asking about explicitly.

Everything else. Shipping is not itemised on H1 listings. Training is not offered as a package. Integration, fixtures, safety engineering, insurance and support are all on the buyer.

Can you still buy an H1, or has it been discontinued?

Neither, quite. It is a legacy SKU in a state of managed decline.

It is still listed on Unitree's official store at $90,000, quote-only. It was still on stage in February 2026, performing parkour, Drunken Fist and nunchaku routines at the Spring Festival Gala alongside the newer H2 doing a sword act. But the main US reseller shows it sold out, the H1-2 has vanished from the official store and survives only through dealers, and Unitree's product messaging has moved entirely to G1, R1, H2 and whatever Superman becomes.

A claim that the H1 reached end of life in March 2026 circulates on robot catalogue sites. We could not verify it — there is no Unitree announcement to that effect, and the product remains listed. Treat it as unconfirmed rather than repeating it as fact.

The practical situation for a 2026 buyer is that you would be acquiring a superseded platform, probably from remaining channel inventory, at three to four times the price of its replacement, with no published warranty and an unresolved US import question. There are good reasons to do that — a lab standardising on hardware that matches published papers is the obvious one — but they need to be good.

Is the Unitree H1 really the fastest humanoid robot?

It was, briefly, in a specific sense, and the claim has been overtaken twice — once by Unitree itself.

The 3.3 m/s figure printed on Unitree's own specification page comes from a video released on 1 March 2024. It is a company claim based on that footage. It is not a Guinness World Record; no Guinness entry names the H1, and outlets that wrote "Guinness" were being loose with the word. The venue of the run has never been disclosed. The certified bipedal 100 m record remains Oregon State's Cassie at 24.73 seconds, set in 2022.

In April 2026 Unitree released video claiming 10.1 m/s peak speed in a 100 m sprint. The caveats matter: that is peak, not average, no total time was released, the robot visibly wandered off a straight line, the test unit was roughly 1.78 m and 62 kg rather than a stock 47 kg H1, and Unitree itself said a measurement error was possible.

Then on 16 August 2026, three days before its Shanghai listing, Unitree unveiled a new machine called Superman claiming a 2 m standing high jump and 12.66 m/s top speed on 0.85 m legs, developed in a little over three months. No test protocol, no payload data, no surface, no repeatability, no independent verification, no hands and no product. It is a prototype timed for an IPO week.

For an H1 owner, none of this changes anything functional. Superman shares no announced hardware, firmware or SDK path with the H1. What it does is confirm that Unitree's R&D attention has left this platform entirely.

What can you build on the H1 — SDK and software

This is the H1's strongest remaining argument, and it is a real one: more published research runs on this robot than on anything else in its class.

The core SDK is unitree_sdk2, C++ under BSD-3-Clause, targeting Ubuntu 20.04 on aarch64 and x86_64. Unitree's own open-source materials describe it as the package used to develop Go2, B2 and H1 robots. ROS 2 support comes via a published driver, and there are official URDF models — which is where the 19-degree-of-freedom count comes from, versus the 18 you get counting five per leg and four per arm on the product page. Both numbers are Unitree's, from different layers, and the discrepancy is benign.

For reinforcement learning there is unitree_rl_gym, which explicitly supports Go2, H1, H1-2 and G1 across Isaac Gym and MuJoCo, with a sim-to-real deployment path — roughly 3,500 stars and nearly 600 forks, which for a robotics repository is a genuine community. Two adjacent projects are commonly confused with it and shouldn't be: unitree_rl_lab is built on Isaac Lab rather than Isaac Gym, and unitree_rl_mjlab targets MuJoCo's newer environment tooling.

The academic record is the payoff. Stanford's HumanPlus full-stack shadowing and imitation work ran on an H1 and published in June 2024. Carnegie Mellon's LeCAR Lab built H2O and OmniH2O on an H1 EDU, published through 2024. UC San Diego's ExBody whole-body control paper states plainly that it adopted the Unitree H1. If your lab wants to reproduce or extend that line of work on identical hardware, that is a coherent reason to buy an H1 in 2026 and essentially the only one.

One caveat we could not resolve: Unitree gates secondary development behind EDU versions across its range — its H2 page states that only the EDU version supports development — and we could not confirm whether a base, non-EDU H1 carries the same restriction. Ask before ordering.

How dangerous is an H1, and what goes wrong?

A 47 kg robot that can run at 3.3 m/s has no fenceless-operation story, and the incident record backs that up.

In February 2025, at a festival in Tianjin, an H1 lunged toward spectators. Unitree described it as an accident caused by a software setting or a malfunctioning sensor. The clip went viral as evidence of robot menace, which it wasn't, and as evidence that these machines need a controlled cell, which it was.

In May 2025 factory footage showed a hoisted H1 thrashing near handlers; that one was never officially explained. In July 2025 a crane-suspended H1 lashed out, and the explanation given was that a full-body control policy was running while the feet weren't touching the ground — the policy operating well outside its training distribution. At the World Humanoid Robot Games in August 2025 the field as a whole crashed extensively, with robots tripping and, in the 1,500 m, reportedly running into people.

The pattern is consistent and reassuring in one narrow sense: these are control policies encountering conditions they were not trained for, not anything emergent or unpredictable in a deeper way. But the practical implication for a buyer is firm. Plan for a fenced or tethered test cell, a hard emergency stop within reach of a human at all times, and a written procedure for what happens when a policy runs outside its envelope. Do not plan to operate an H1 near people who have not been briefed.

What are the security risks of owning a Unitree H1?

This is the section that has changed most since our last update, and unlike most Unitree models, the H1 is named directly in the older and more damaging disclosure.

In September 2025, researchers Andreas Makris and Kevin Finisterre, with Konstantin Severov, published UniPwn — four vulnerabilities tracked as CVE-2025-35027, CVE-2025-60017, CVE-2025-60250 and CVE-2025-60251, affecting the Unitree Go2, G1, H1 and B2 on all firmware current at the time. The mechanism was as bad as it sounds: a hardcoded AES key shared across every device, authentication satisfied by encrypting the literal string "unitree", and then unsanitised command injection through the Bluetooth Wi-Fi provisioning fields straight into system calls in a Wi-Fi restart script. Root, over Bluetooth, with no user interaction. It is wormable — a compromised robot can scan for others in Bluetooth range and infect them.

Unitree's handling was poor. Disclosure attempts began in May 2025, the company said fixes would require quarters or years, removed the researchers from email threads, and stopped responding in late July 2025. On 29 September 2025 it said publicly it was aware and had completed the majority of fixes. No H1-specific fix has been publicly announced since.

Then on 27 August 2026 Olivier Laflamme published two further root-level flaws, CVE-2026-76639 and CVE-2026-76640, the second exploitable over Bluetooth without pairing by chaining a permissive GATT characteristic, a cloud endpoint that returned device keys to accounts that did not own the robot, and a buffer overflow in the Bluetooth server. Those were confirmed on the G1 only. The researcher explicitly declined to claim that shared software stack proves other models are affected, and the H1 is not named in either CVE. But the attack surface is the same Bluetooth Wi-Fi provisioning subsystem the 2025 flaws lived in, so assume shared-architecture risk rather than safety.

Separately, security reporting from October 2025 described Unitree units transmitting data to servers in China at roughly five-minute intervals, naming the Go2, B2, G1 and H1 as affected by the underlying vulnerability set. The destination host and payload have not been characterised publicly. The Pentagon has also added Unitree to a list of firms it says have ties to or aid the Chinese military.

The practical position: treat an H1 as a hostile device on your network. Isolated VLAN, no route to anything sensitive, Bluetooth radios disabled outside provisioning, and physical control of who gets within Bluetooth range of it. If your institution has an information security review process, this robot needs to go through it.

Can you legally import a Unitree H1 into the US?

For an existing unit, yes. For a new import in 2026, this needs a real answer before you spend the money.

On 28 July 2026 the FCC released Public Notice DA 26-786, adding foreign-produced advanced robotic devices to its Covered List. The effect is that covered new or previously unauthorised robot models cannot receive FCC equipment authorisation, which is required to import, market or sell a radio-equipped device in the US. Scope is mobile ground robots over about 2 kg with environmental sensing, network connectivity and control software — humanoids, quadrupeds and warehouse AMRs. Fixed industrial arms are excluded. The test is country-neutral: "foreign-produced" means failing the Buy American domestic-end-product standard of US manufacture with more than 65% domestic component cost, so European, Japanese and Korean robots sit inside the same wording.

The rule is prospective and existing owners are explicitly fine. It is not a customs ban and not a recall. Machines with FCC authorisation before 28 July 2026 remain legal to import, market, sell, own and resell, no authorisations were revoked, and there is no restriction on continued use of units already in the country. A companion order also guarantees that authorised devices can keep receiving software and firmware updates through at least 1 January 2029, which matters given the outstanding security work above.

Here is the H1-specific problem. Unitree cleared its current lineup in the weeks before the deadline. Published grant dates under its grantee code cover the G1 in March 2025, Go2 in May 2024, B2 in October 2024, the R1 on 22 June 2026, and the H2 and A2 quadruped both on 30 June 2026. No separate authorisation is visible for the H1. That is suggestive but not conclusive: absence from a public database is not proof, and filings can sit under a different legal entity. Treat the H1's status as unresolved, and put the burden on the seller to show you the grant.

So: read the FCC ID off the physical label of any H1 you are offered and check it before payment. If a dealer cannot produce a grant, you are being asked to import a robot whose legal status you cannot establish. On top of that sits a reported 25% Section 301 duty on Chinese robots, which is the stable part of the import cost. Budget for a compliance review, not just for shipping.

Unitree itself has publicly dismissed the impact of the FCC rule, pointing to its authorised G1, H2, R1, Go2, B2 and A2. Note which model is missing from that list.

Where have H1 units actually been used?

In labs and on stages, and the stage appearances are more often misattributed than any other fact about this robot.

Academically, the H1 has a strong and verifiable record: Stanford's HumanPlus in June 2024, Carnegie Mellon's LeCAR Lab H2O and OmniH2O work on an H1 EDU across 2024, and UC San Diego's ExBody expressive whole-body control paper in February 2024. Beyond published papers, no dated list of institutional purchasers exists — Unitree does not publish customers and procurement records are not indexed — so treat any confident list of universities owning H1s with suspicion.

On stage: the 16 robots that performed Zhang Yimou's Yangge dance segment at the CCTV Spring Festival Gala on 28-29 January 2025 were H1s, not G1s. That is the single most misattributed appearance in Chinese robotics. At the 2026 Gala on 16 February, the H1 performed parkour, Drunken Fist and nunchaku while the newer H2 did a sword routine.

In competition: at the World Humanoid Robot Games in Beijing in August 2025 the H1 won gold in both the 1,500 m and the 400 m, with Unitree taking four golds and eleven medals overall. The 1,500 m time is reported inconsistently across outlets — figures from 6:29 to 6:34 circulate — so cite it as approximately six and a half minutes rather than picking one.

Two corrections worth carrying: the Iron Fist King boxing event of May 2025 used G1s, not H1s, as Unitree's own event page makes clear. And the H1 did not win the Beijing humanoid half-marathon — the 2025 winner was Tiangong Ultra, and the 2026 winner was a machine from a Shenzhen firm.

Commercially, there are no disclosed H1 deployments. Unitree's own prospectus supports that: industrial use is around 9% of its humanoid revenue and that is overwhelmingly G1-class hardware.

H1 vs H2, G1 and the wider field

Within Unitree's own range the H1 is now the worst value in the catalogue, and it is not close.

The H2 at $29,900 is taller-equal, has 31 degrees of freedom against the H1's 18 to 19, 360 N·m joint torque, 2,070 TOPS of onboard compute, a bionic face and over-the-air updates. It also holds an FCC grant dated 30 June 2026. US dealers price it at $40,900 commercial and $68,900 for the EDU version, which is still below the H1's dealer pricing. The G1 at $13,500 covers most research needs at half the H1's degrees of freedom but a sixth of the price. The H2 Plus at $100,000 is the flagship.

The H1-2 is a different trade rather than an upgrade: 178 cm, 70 kg, 27 degrees of freedom with seven per arm, a 7 to 21 kg payload and roughly two hours of runtime — but under 2 m/s. It swaps the H1's speed for arms and carrying capacity, and it is sold only through dealers now, around $122,000.

Outside Unitree at similar money: Fourier's GR-2 is reported around $125,000 and the GR-3C around $85,000, both from catalogue sources rather than the manufacturer. Booster's T1 is quoted around $34,000 by distributors. Agility's Digit is not sold outright to most buyers — the model is a lease at roughly $8,500 a month. Apptronik's Apollo has no public price at all, though the company raised $520 million at a $5 billion valuation in February 2026 and runs pilots at Mercedes-Benz and GXO. Figure, Atlas and Optimus publish nothing.

So the H1's competitive position is: outclassed inside its own product line at a third of the price, priced above better-specified alternatives outside it, and carrying a regulatory question none of them have.

Is Unitree a safe company to buy an obsolete robot from?

The company has never been stronger financially, which cuts both ways for an H1 buyer.

Unitree listed on the Shanghai STAR Market on 19 August 2026 under ticker 688836, priced at ¥150.80 a share for a valuation near ¥61 billion — about $9 billion — raising roughly ¥6.1 billion. Retail demand oversubscribed the offering more than 8,000 times, a STAR Market record. On debut the stock peaked at ¥1,100 intraday, up about 629%, and closed at ¥845, up 460%, briefly valuing the company around $66 billion. It then slumped: by 25 August it was roughly 45% below the peak, and it closed 28 August at ¥585 with a market capitalisation near ¥237 billion, about $33 billion.

The prospectus numbers explain both the enthusiasm and the correction. Revenue went from ¥159 million in 2023 to ¥393 million in 2024 to ¥1.70 billion in 2025, with 2025 net profit of ¥278 million and gross margin around 60%. Humanoid units sold went from 5 in 2023 to 412 in 2024 to 5,215 in 2025. Overseas revenue was 43.65% of main-business revenue in 2025. Founder Wang Xingxing retains around 20%, which made him China's first humanoid robot billionaire several times over.

The number that matters for an H1 buyer is the average selling price, which collapsed from around $85,000 in 2023 to about $25,000 over the first nine months of 2025. That is the H1-to-G1 mix shift in a single figure, and it tells you exactly where Unitree's volume, engineering and support attention now sits. R&D also fell from 31.4% of revenue in 2023 to 7.7% over nine months of 2025.

Translation for a buyer: Unitree is not going anywhere, but the H1 is a rounding error in its business. Do not expect firmware attention, parts priority or a security fix schedule for a platform that represents a tiny fraction of a company shipping thousands of $13,500 robots a year.

Should you buy a Unitree H1 in 2026?

For most people, no, and the reason is not the robot — it's the arithmetic.

Buy one only if you are a research group deliberately standardising on the exact hardware behind the HumanPlus, OmniH2O or ExBody line of work, you can verify the FCC grant on the physical unit before payment, you have an information security process that can accept a device with unpatched wormable Bluetooth vulnerabilities from 2025, and you have a fenced test cell with a hard emergency stop. Under those four conditions the H1 is a defensible, even sensible, purchase, and the published research base behind it is genuinely unmatched.

Don't buy one if you want the most capable Unitree humanoid: that is the H2 at $29,900, with more degrees of freedom, far more compute, over-the-air updates and a clean FCC grant. Don't buy one if you want value: the G1 at $13,500 covers most research at a sixth of the price. Don't buy one if you want manipulation: the base arms have four degrees of freedom and no hands, and the H1-2 exists precisely because the H1 was the wrong shape for that work. Don't buy one expecting household or commercial tasks; there is no vendor claim, no demonstration and no included stack, and Unitree's own revenue split confirms nobody is doing it.

And don't buy one because it is the fastest humanoid in the world. That claim was defensible for a full-size humanoid in 2024, and Unitree has since beaten it twice with other hardware. It is a lovely piece of engineering with a real place in the history of this field. It is just not a 2026 purchase for most of the people searching for it.