The Unitree G1 humanoid robot review nobody has written

The Unitree G1 humanoid robot review you are looking for does not exist. Nobody has lived with one for six months and published what broke. I have not touched one either, and I am not going to pretend otherwise.

What does exist is a paper trail, because the people who own G1s are mostly institutions, and institutions file papers and bug reports rather than reviews. Everything below is their account, linked and dated. None of it is mine.

Who owns one

Unitree's own IPO filing settles the question. In The Robot Report's March 25, 2026 read of the prospectus, ahead of the company's STAR Market listing, 73.6 percent of humanoid revenue in the first nine months of 2025 came from research and education. Another 17.4 percent came from commercial uses such as demonstrations and display environments. Just 9.01 percent was industrial.

Same filing: 3,701 humanoids produced and 3,551 sold in those nine months, with average selling price down from about $85,000 in 2023 to about $25,000.

There is no household column. Not a small one, not a rounding error. The category most buyers imagine themselves in does not appear in the maker's own revenue breakdown.

What owners report when it works

An arXiv search for the exact phrase "Unitree G1" returned 166 papers. That is the real review corpus, and it is a good one, because a paper has to state what the robot did and let other people try to repeat it.

Two examples from opposite ends of the range. Shooting for Contact (Esteban et al., submitted 2026-08-04) reports "zero-shot sim-to-real transfer on the Unitree G1 through command-conditioned contact-rich crawling and a highly dynamic 180-degree jump-turn." So: it crawls, and it can whip 180 degrees mid-jump, first try, from a policy trained in simulation.

HumanoidTurk (Lee, Suzuki and Hong, submitted 2026-01-26) is the closest thing to a consumer report in the pile. The team parked a G1 behind a VR driving simulator's chair and had it move the chair in time with in-game g-forces, then ran sixteen participants through it. Their finding, verbatim: "humanoid feedback enhanced immersion, realism, and enjoyment, while introducing moderate costs in terms of comfort and simulation sickness." The interviews landed on something no spec sheet prints, the robot's "consistency and predictability in contrast to the adaptability of human feedback."

One caveat on all 166 of those. I read the abstracts of the twenty-five most recent by hand, and not one mentions a hardware failure, a thermal limit, or downtime. Papers report the run that worked. For the runs that did not, you have to go where the complaints get filed.

What owners report when it does not work

Unitree publishes its reinforcement-learning training code on GitHub, and the issue queue under it is the most useful G1 owner forum in existence. These are people who paid, unpacked, and hit a wall.

Filed	Issue	What the owner reports	State
2025-02-26	#43	Joint order in the G1 config appears not to match what the URDF parses at runtime	Open
2025-06-05	#64	Policy deployed on a Jetson Orin NX performs badly: joint instability, loss of balance	Open
2025-07-30	#70	Legs flying around during training, on an unmodified default config	Open
2026-01-13	#83	The real-robot deploy script launches, then the terminal freezes and the robot does nothing	Open
2026-03-04	#89	Motor chattering and vibration when running Unitree's own pretrained walking policy; their self-trained policy does not do it	Open
2026-04-16	#94, #95	Joint ordering, observation format, scaling and command ranges for the 23-joint G1 policy are not documented anywhere	Open

Over in Unitree's C++ SDK, a pull request dated 2026-04-29 fixes a CRC mismatch caused by struct padding, in which the firmware silently drops the command frame and the motors never come under the user's control at all. A buyer found that and wrote the patch.

Read the pattern rather than any single row. The hardware works. The documentation around it is where owners lose their week: undocumented joint order, undocumented scaling, timing that falls apart on the onboard computer, and a walking policy from the factory that makes the motors buzz.

The repo that stopped moving

The same repository, via the GitHub API: 3,480 stars, 576 forks, 59 open issues, last commit 2025-07-25.

More than a year without a push, while the issues kept arriving. That is the shape of the support here: the code you are meant to build on is public, popular, forked 576 times, and standing still. Check the repo before the crate lands, not after.

The owners who report nothing

Everything above comes from EDU owners, because they are the only ones who can file it. The base G1 at $13,500 "does not support secondary development" per Unitree's own store page (still $13,500 and still backordered), so a base-unit buyer has no code to debug and nothing to post.

What a base G1 leaves behind instead is footage. Boxing matches in Hangzhou, the Beijing robot games, and a customized unit doing marketing appearances in Warsaw, all logged on the G1 tracker page. The most instructive one is the 2025 Beijing half marathon, where the widely shared clip of a G1 going down near the start line came from a machine Unitree said it had not entered: those were customer units, bought by independent teams and run on the teams' own software. An owner report, filmed by everybody, explained by nobody.

What still is not known

Five things I could not source, and will not guess at:

• Reliability. No failure rate, no service-interval data, no published mean time between failures from anyone.
• Repairs. The spec table at unitree.com gives the base G1 an 8 month warranty (18 on the EDU), with no described US service network. What a post-warranty joint actuator costs to replace, and how long it takes, is not documented publicly.
• Real runtime. The spec is "about 2h." I have found no independent measurement of what that becomes with a payload on a working day.
• The EDU price. Still quote-only. Resellers listed $43,900 to $73,900 as of 2026-07-25, which is a range, not a price.
• Private owners. I could not verify a single long-term review by an individual who bought one with their own money. Threads and clips exist; identity and ownership behind them do not check out from here, so none of it is quoted above.

How to meet one before you spend $13,500

Cheapest route first: rent someone else's. The Construct's humanoid masterclass, posted to the ROS community on 2026-07-23, offers participants the chance to "remotely access and experiment with a real Unitree G1 humanoid robot hosted at The Construct's robotics laboratory." The same board has carried G1 certificate-training listings in Barcelona since 2025. Neither route is free. Neither one asks you to clear 35 kg of robot through customs first.

If the purchase is for a lab or a classroom, the platform comparison in humanoids for education and research covers what else that budget reaches. If it is for a home, read the price history and then the import situation for Chinese humanoids, which changed on July 28, 2026. And if $13,500 is more than an experiment justifies, the same maker's R1 starts at $4,900 with the same secondary-development caveat.

If you are close to buying one, read in the order the record suggests: the G1 tracker entry for the current price and stock, then the open issues above for what your first month is likely to look like. Getting the crate through the door is the easy part.