A robot emergency stop (E-stop) is a safety function that halts a robot's hazardous motion on demand, overriding every other command. On a humanoid, the E-stop also has to keep the robot from falling, and from a security point of view it is the one control an attacker must never be able to block, fake or switch off. This entry explains how a robot emergency stop works, why humanoids make it harder, and how a kill switch can be attacked.
Last updated: November 19, 2026
How a robot emergency stop works
Two machinery standards define most E-stops. ISO 13850 sets the rules for the function itself: as summarized by drive maker Kollmorgen, the E-stop "must have priority over all other functions and controls in all operating modes", must be available at all times, and resetting it must not restart the machine.
IEC 60204-1 defines how the machine stops:
| Stop category | What happens | Used for E-stop? |
|---|---|---|
| Category 0 | Power to the motors is cut immediately (uncontrolled stop) | Yes |
| Category 1 | Motors brake the machine to a controlled stop, then power is cut | Yes |
| Category 2 | Controlled stop with power kept on | No |
Which category to use comes from a risk assessment of the machine. A related term, a kill switch, usually means the same thing in consumer and AI contexts, though it is sometimes used loosely for a software command that only pauses a program.
Why the emergency stop matters for humanoid robots
On a fixed industrial arm, cutting power (Category 0) is usually the safest move. On a walking humanoid it is not. A March 2026 paper from Carnegie Mellon University and Siemens, Learning Safe-Stoppability Monitors for Humanoid Robots, notes that because humanoids rely on balance and ground contact, "abrupt power loss inevitably induces catastrophic falls". The authors argue a humanoid E-stop must run a fallback controller that keeps balance while it brings the robot to rest, and tested that approach on a Unitree G1.
The draft humanoid safety standard ISO 25785-1 points the same way. According to an i-SCOOP summary of the draft, it favors controlled descent over collapse and defined stability behavior when an E-stop is triggered. Some vendors already do this: Humanoid.guide describes Agility's Digit dropping to its hands and knees before powering down. Our humanoid robot security standards guide covers where ISO 25785-1 and ISO 10218 stand.
That is a security problem too (analysis): the smarter the stop, the more software sits in the stopping path, and software can be attacked.
How an E-stop can be attacked
- Blocking the stop. In a November 2026 sponsored article in IEEE Spectrum, security firm VicOne said its researchers "exploited a vulnerability to prevent a robot's remote-stop command from working". The robot and software were not named, and no CVE has been published.
- Abusing the stop. In 2019, Trend Micro researchers showed that industrial radio remote controllers for cranes and hoists could be held in a persistent denial-of-service state "through emergency stop (e-stop) abuse", because none of the products they tested used rolling codes.
- Bypassing the safety layer. If stops and limits live only in software that also runs the AI, a compromise of that software can reach them. That is why our humanoid robot attack surface map puts an independent, hardwired E-stop beneath every other layer.
Example
A warehouse runs humanoids with a wireless E-stop on each supervisor's tablet. A secure design keeps a hardwired stop button on the robot and a separate safety-rated radio channel, so a hacked tablet app or Wi-Fi outage can neither block a stop nor freeze the fleet with fake ones. The robot's controlled-descent routine runs on a safety controller, not the main AI computer.
Related terms
- Physical AI
- Vision-Language-Action (VLA) model
- ROS 2 security
- Safety filters, and how they can be turned into collision tools
FAQ
What is the difference between an emergency stop and a kill switch? In robotics they usually mean the same thing: a control that halts hazardous motion and overrides everything else. "Kill switch" is also used for software commands that only pause a program, which do not meet the E-stop standards.
Why can't a humanoid robot just cut power in an emergency? Cutting power to a walking humanoid makes it fall, which can injure people and damage the robot. Humanoid E-stops increasingly use a controlled stop that keeps balance, then lowers the robot before power is removed.
Can a hacker disable a robot's emergency stop? If the stop depends on the same software or network an attacker controls, yes. A hardwired E-stop on a safety-rated controller, independent of the robot's AI and apps, is the main defense. Check published flaws in our Humanoid Robot Vulnerability Tracker and test your setup with the free Humanoid Robot Security Checklist.
Get the weekly Humanoid Threats Brief → Subscribe


