Standards
Humanoids leave the cage: why 'power off' isn't a safe state for physical AI that balances
What's the safe state of a machine that has to balance?
For an industrial arm, the answer is old and simple. Cut the power. The arm stops. That's Stop Category 0, and a whole family of machine safety standards leans on it.
Cut the power to a walking humanoid and it falls. For physical AI that balances, "power off" becomes a hazard of its own. A preprint from 3 August 2026 gives this a name: the fail-passive gap.
The authors built a certified external safety chain around a Unitree G1. Light curtain, emergency stop, fail-safe PLC and wireless PROFIsafe. Established methods can quantify each part. That's the point of the exercise. It pins the uncertifiable piece to the robot side. The certifiable reaction in their reference is contactor-based power removal. The paper calls it "exactly the element a balancing humanoid cannot have". The authors also say they "deliberately do not claim" end-to-end PL e or SIL 3.
Their test bed is one pick-and-place cell, 3 m by 1.5 m and semi-enclosed. They write that the safe state of a legged robot is "an actively-controlled state". The robot has to keep working to stay safe.
Meanwhile the cages are coming down. Forbes reported on 2 October that Agility's Digit 5 is built to work right beside people with no safety cage. The same piece says Digit 5 gets collision avoidance in 2027, according to Agility. Boston Dynamics says its production Atlas supports "fenceless guarding".
On 13 October, Korea moved. Its Ministry of Trade and Industry granted the robot maker Robros a demonstration exemption. 20 bipedal humanoids will work in spaces shared with or next to people. The sites are four companies: Nepes, Sungwoo Hitech, HD Hyundai Samho and Delivery Lab. The conditions are risk assessments and safety measures. The government will use the data to draw up future safety standards.
So where's the standard? ISO 25785-1 covers robots that "could become unstable in the absence of power". On 29 October the ISO catalogue listed it as a committee draft at stage 30.60, close of comment period. Its scope is industrial only: workplaces that keep the public out or restrict its access. Homes are out. The draft points to ISO 13482 for those.
| Question | Where it stands |
|---|---|
| Safe state when power drops | The robot falls |
| External chain (curtain, e-stop, PLC) | Certifiable with known methods |
| Robot-side balancing reaction | The gap. Not certified |
| ISO 25785-1 | Committee draft, stage 30.60, industrial only |
| Korea | 20 humanoids, uncaged, on conditions |
Last week the field met in Detroit. A3's announcement for its International Robot Safety Conference (3 to 5 November) put Agility and Boston Dynamics in one session. Its topic was safety challenges for dynamically stable industrial mobile robots. That's the draft standard's own subject line.
Here's the honest limit. The fail-passive paper is a preprint. It's one feasibility study on one robot in one cell. It doesn't prove that nobody can certify a balancing humanoid. It shows where today's certified methods stop. I read the conference announcement, not the session itself. And the ISO stage is the one I saw on 29 October.
Bottom line: an uncaged humanoid needs a safe state that keeps it under control. No published standard defines that state yet.
Safety engineers: what's the safe state in your humanoid's risk assessment, and who proved it?
Part 1 of 12 · The Survey: what keeps physical AI safe in 2026.
Sources
- arXiv — Ding, Cui, Wang, Wen, Toward Certified Functional Safety for Industrial Humanoid Robots: The Fail-Passive Gap and a Feasibility Study, 2608.02809, preprint (read 29 October 2026)
- ISO — ISO/CD 25785-1, dynamically stable industrial mobile robots, Part 1: Robots, catalogue entry 91469 (read 29 October 2026)
- Seoul Economic Daily — Korea Clears Safety Rules to Put 20 Humanoid Robots on Factory Floors, 14 October 2026 (read 29 October 2026)
- Forbes — Humanoid Robots Are Leaving Safety Cages. Proving Actual Safety Is Harder, 2 October 2026 (read 29 October 2026)
- A3 — International Robot Safety Conference comes to Detroit, November 3-5, 16 October 2026 (read 29 October 2026)
Cite this
APA
Melchior, M. (2026, November 10). Humanoids leave the cage: why 'power off' isn't a safe state for physical AI that balances. Physical AI Safety. https://physical-ai-safety.com/blog/humanoids-leave-the-cage-fail-passive-gap
BibTeX
@misc{melchior2026humanoids,
author = {Melchior, Mati},
title = {{Humanoids leave the cage: why 'power off' isn't a safe state for physical AI that balances}},
howpublished = {Physical AI Safety},
year = {2026},
month = nov,
date = {2026-11-10},
url = {https://physical-ai-safety.com/blog/humanoids-leave-the-cage-fail-passive-gap}
}- Published