# Humanoid at home: no dedicated standard found, researchers describe data being sent without warning

> We found no standard specific to the household humanoid, the only full benchmark we found covers a G1, and researchers have described data being sent out without warning. What Fraunhofer IPA, Unitree, 1X and France’s data regulator CNIL say, according to the pages read on 31 October 2026.

Published on 2026-10-31 · Updated on 2026-11-08

## No standard specific to the household humanoid

A humanoid the size of a G1, weighing about 35 kg, can fall or hit someone. Yet we found no standardized answer to “is it safe in my home?” The draft standard ISO 25785-1 covers industrial mobile robots with actively controlled stability (legged, wheeled or other). It is at committee draft stage, and a national standards body’s listing shows the voting and comment period closing on 8 July 2026. Its scope is industrial: it excludes robots used outside industry, an industrial setting being defined as a workplace from which the public is excluded or where access is restricted. The listing points to other standards, such as ISO 13482, for mobile robots outside industry; we have not checked how they apply to humanoids. Fraunhofer IPA writes separately that safety standards for humanoids are not expected before 2028, naming this same ISO 25785-1 project.

## The only full benchmark we found, on a G1

On 27 May 2026, Fraunhofer IPA published the first application of its six-category humanoid benchmark, on an [Unitree G1](/robots/unitree-g1) EDU-4 fitted with three-finger Dex3-1 hands, delivered in May 2025 and running firmware 1.04. First safety result: collision forces above 500 N can occur, far above the pain thresholds permitted by the standard, according to the institute, which gives neither the threshold value nor, in that sentence, the name of the standard (the release elsewhere cites ISO 10218 and ISO TS 15066 for functional safety). The measurements used the same force sensors as for collaborative industrial robots. Second result: a critical Bluetooth vulnerability, present in the software at the time of the test, allowed complete remote control of the robot. The institute writes that it has since been fixed, without saying in which version. On the positive side, stabilization on a ramp is rated good.

The test has limits. The release describes a G1 EDU-4, from the EDU range that Unitree’s page opens to secondary development, without saying how many units were tested. It mentions industrial applications (semiconductors, pharmaceuticals, food) and no test in a home. It gives no walking speed, gripping force or load capacity. The maximum run times measured on one charge, 2 h 49 min standing still and 1 h 49 min in a typical scenario mixing standing and walking, deserve the same caution.

## What the makers say

Unitree’s G1 page asks users to keep a sufficient safe distance between the robot and people and to use it with caution. It states that some sample functions are still being developed and tested, that the humanoid industry is at an early stage, and that individual users are strongly advised to understand the robots’ limitations before buying.

For [NEO](/robots/1x-neo), 1X describes a soft body of 3D lattice polymer structures, gentle movements that it says make the robot safe around people, and a noise level of 22 dB, quieter, the company says, than a modern refrigerator. These are the maker’s claims: we found no independent measurement of NEO. Road to VR (30 October 2025) reports, citing 1X’s chief executive, that NEO cannot pick up items that are too heavy, hot or sharp.

## Privacy: what leaves the robot

Three researchers, Víctor Mayoral-Vilches (Alias Robotics), Andreas Makris (thinkAwesome! GmbH) and Kevin Finisterre (affiliation not published), put online a security assessment of a G1 in September 2025, along with a second paper by Mayoral-Vilches alone. They report that the robot sends sensor and state data to two servers every 300 seconds without notifying the operator, and argue that this breaches Articles 6 and 13 of the GDPR. They also describe a flaw in the Bluetooth setup protocol that gives root access, exploitable with AES keys that are the same on every unit, and encryption built on static keys. Decrypt (22 September 2025) reports, about Mayoral-Vilches’s report, that it finds no privacy policy, no disclosure of the collection, no way to refuse it, and no visible or audible signal when the robot records or transmits; we could not verify them word for word in the arXiv papers read. The three-author paper points to missing consent or transparency mechanisms, and states that its limited inspection of the telemetry did not let the authors confirm audio capture without a user indicator. The site writes that Unitree had not responded immediately to its request for comment.

Read with some distance: these are two arXiv papers by authors linked to one another, on a robot whose firmware version is not given, and the pages we read mention neither a published response from Unitree nor an independent replication. Jan Liphardt, chief executive of OpenMind, quoted by Decrypt, judges that the report veered toward alarmism and that the architecture described is fairly typical. The G1 page today links to a “Security and Privacy” section whose text we did not read. Finally, nothing says that the Bluetooth flaw these researchers describe is the one Fraunhofer reported.

## At 1X, a human can look into your home

According to Engadget (29 October 2025), reporting on a Wall Street Journal interview with 1X chief executive Bernt Børnich, buyers of a NEO delivered in 2026 must agree that a human operator will see inside their home through the robot’s camera. The owner schedules the sessions and sets no-go zones; according to Engadget people can be blurred (Road to VR writes only that NEO could “potentially” do so), and the operator cannot take control without the owner’s approval. Børnich explains that the robot needs real-world experience: without your data, he is reported to have said, the product cannot be improved (paraphrased). According to Road to VR, Børnich frames privacy as a sliding scale of permissions: the more data you give up, the better NEO works.

NEO’s page on 1x.tech, read on 31 October 2026, says the owner can schedule a 1X expert to guide the robot through an unknown task. It covers neither blurring, nor no-go zones, nor data use: we know these settings only from the press. It links to a “Privacy & Cookies” policy that we did not read. It announces deliveries mainly in the United States in 2026, then other markets from 2027, with no date for France.

## The CNIL’s advice on connected robots

The CNIL page on connected robots dates from 26 November 2019 and is about connected kitchen robots, not humanoids; applying it to a humanoid is our reading. Its three pieces of advice are general: learn about the features before buying, match the purchase to real needs (is the connectivity worth the extra cost and the risk to your data?), and check that unused functions can be turned off, physically (for example with a button on the device) or in software, notably Wi-Fi and the microphone. For a robot that sees and hears the home, these three habits apply all the more.

## Questions to ask before buying

Does the seller state a safety distance and the maximum impact force, and who measured it? Where does the sensor data go, how often, and does a signal warn when the robot is recording? Can the camera, the microphone and the connection be switched off? Can a human look into your home, under what conditions, and can you refuse? Until when does the robot receive security updates? Is it delivered and covered by a warranty in France? Better to get these answers in writing before paying.

## Sources

- [Fraunhofer IPA, press release of 27 May 2026, read on 31 October 2026](https://www.ipa.fraunhofer.de/en/press-media/press_releases/benchmark-for-humanoid-robots.html)
- [Institute for Standardization of Serbia, listing of draft ISO/CD 25785-1, read on 31 October 2026](https://iss.rs/en/project/show/iso:proj:91469)
- [Unitree, G1 page, read on 31 October 2026](https://www.unitree.com/g1)
- [arXiv, “Cybersecurity AI: Humanoid Robots as Attack Vectors” (Mayoral-Vilches, Makris, Finisterre), 17 September 2025, read on 31 October 2026](https://arxiv.org/abs/2509.14139)
- [arXiv, “The Cybersecurity of a Humanoid Robot” (Mayoral-Vilches), 17 September 2025, read on 31 October 2026](https://arxiv.org/abs/2509.14096)
- [Decrypt, article by Jason Nelson of 22 September 2025, read on 31 October 2026](https://decrypt.co/340762/humanoid-robot-cheaper-rivals-full-security-flaws-researchers)
- [Engadget, article by Mariella Moon of 29 October 2025, read on 31 October 2026](https://www.engadget.com/ai/1x-neo-is-a-20000-home-robot-that-will-learn-chores-via-teleoperation-040252200.html)
- [Road to VR, article by Scott Hayden of 30 October 2025, read on 31 October 2026](https://roadtovr.com/helper-robot-neo-vr-telepresence/)
- [1X, NEO page, read on 31 October 2026](https://www.1x.tech/discover/neo-home-robot)
- [CNIL, “Robots connectés et données personnelles : les conseils de la CNIL” (in French), page of 26 November 2019, read on 31 October 2026](https://cnil.fr/fr/robots-connectes-et-donnees-personnelles-les-conseils-de-la-cnil)

---

HTML version: https://vuntum.com/en/news/a-humanoid-at-home-what-safety-and-privacy-tests-say

Factual data: CC BY-NC 4.0. Attribution: “Source: Vuntum (https://vuntum.com)” with a link to the page quoted. Editorial text: all rights reserved, short quotation with attribution.
