E-stops and safety-rated stops on humanoid robots: what ISO 13849 means
Cutting power makes a factory arm safe but makes a walking humanoid fall. Here is how stop functions are rated under ISO 13849, and why legged robots are a special case.
Von Redaktion von humanoid4rent · veröffentlicht 15. November 2026
On a fixed industrial machine, the safest response to an emergency is usually to cut power. On a walking humanoid, doing so can make the robot fall, possibly onto someone. Emergency and protective stops are therefore designed differently for legged robots, and the usual way of rating them under ISO 13849-1 does not yet fit them well. Below are both points and what to ask a supplier.
What a safety function is
A safety function is a specific job a machine's control system does to reduce a risk. Examples are "stop when the emergency button is pressed", "slow down when a person enters a zone" or "never exceed a set speed". The function is a whole chain, not one part. It runs from the button or sensor, through the logic that decides what to do, to the motors and brakes that carry it out.
Most robots also have plenty of ordinary software that avoids obstacles or keeps a safe distance. That software can work very well, but it is not automatically a safety function in the formal sense. The difference is whether the chain has been designed, analysed and documented to a recognised standard, so that its chance of failing dangerously is known and limited.
ISO 13849-1 and performance levels, in plain terms
ISO 13849-1 is the standard most often used in Europe to design and assess the safety-related parts of machine control systems. The current edition is from 2023, and the German Social Accident Insurance institute IFA has updated its free SISTEMA calculation software to support it.
The standard grades each safety function with a Performance Level (PL) from a to e. PL a is the lowest and PL e the highest. As IFA's guide to the standard explains, the level a function reaches depends on several factors working together:
- Architecture (Category B to 4): whether the chain has a single channel or redundant channels that check each other.
- Component reliability (MTTFd): how long, on average, parts run before they fail dangerously.
- Diagnostic coverage (DCavg): how many dangerous faults the system detects by itself.
- Common-cause failure measures: protection against one event, such as heat or a power surge, disabling both channels at once.
- Software and systematic measures: how carefully the safety software is specified, tested and kept under change control.
The designer first carries out a risk assessment to set a required level (PLr) for each function. The more severe and frequent a potential injury, and the harder it is to avoid, the higher the required level. The built function must then reach at least that level.
For a busy professional, a useful shorthand is that "PL d, Category 3" means the function was built with redundancy and self-checking. It does not mean the whole robot is "safe".
Why a walking robot cannot just cut power
The classic stop on a fixed machine is what IEC 60204-1 calls stop category 0: power is removed from the motors straight away, usually through contactors. Other categories in that standard bring the machine to a controlled stop first. On a robot arm bolted to the floor, losing power leaves the arm where it is, held by brakes. The machine is "fail-passive": when things go wrong, doing nothing is safe.
A standing or walking humanoid is different. It keeps its balance by constantly adjusting its joints (see our explainer on how humanoid robots balance). Take away the power and it collapses.
In an August 2026 preprint, Caiwu Ding and colleagues call this the "fail-passive gap". They argue that removing power from a walking biped is itself a hazard, which clashes with the assumptions behind the classic stop. They tested their approach on a Unitree G1 EDU (the G1 EDU is the research version of the G1) in a pick-and-place cell.
The authors state clearly that they do not claim full certification at the highest levels, PL e or SIL 3. The paper is a preprint and has not yet been peer-reviewed.
The practical result is that a humanoid's safest stop usually needs the robot to keep working for a few more seconds. It has to stop walking, settle its weight, and ideally lower itself to a stable posture before power is removed. That means the motors, sensors and balance software all become part of the safety chain. Rating a chain that includes a learned balance controller is exactly the part that existing methods struggle with.
How manufacturers handle stopping today
Published approaches show the trade-off clearly.
- Boston Dynamics Spot. Spot's E-Stop service documentation describes two stop levels. CUT removes actuator power at once. SETTLE_THEN_CUT makes the robot try to stop, sit down and then cut power. The service runs on a heartbeat. If a registered controller misses its timeout, for example because the radio link dropped, the robot falls back to CUT.
- Agility Robotics Digit 5. According to IEEE Spectrum's report from September 2026, Agility's approach is that Digit avoids people, stops, or sits down. As a person approaches, the robot puts down what it is carrying and, if needed, sits stably on the ground before they get close. The aim is that the robot is already in a stable posture before anyone is close enough to be hurt by a fall. The same report lists Digit 5 at 1.8 metres and 129 kilograms, which shows why a fall matters.
Spot's documentation describes a software and communications protocol, not a certified, PL-rated safety function. Both examples avoid the problem of cutting power mid-stride by first moving the robot into a state where losing power is harmless.
Making stops predictable: current research
If a robot can only stop safely from some states, it helps to know at every moment whether it is in one of them. A March 2026 preprint by Yifan Sun and colleagues, Learning Safe-Stoppability Monitors for Humanoid Robots, treats this as a question of "stoppability".
The authors train a neural network in simulation, then transfer it to real hardware, to predict whether the robot can carry out its fallback stop sequence safely from its current state. They present this as a step towards certifying fail-safe behaviour. Like the Ding paper, it is a preprint, and it is research rather than a feature shipping in rental robots.
Standards bodies are working on the same gap. ISO is drafting a safety standard specifically for dynamically stable legged robots, covered in our article on ISO 25785-1. Until such work is published, suppliers have no single agreed method for rating a humanoid's stop.
What the labels do and do not tell you
Three cautions help when reading a datasheet or a supplier's answer.
- A PL applies to one function, not the robot. "PL d emergency stop" says nothing about the robot's obstacle avoidance or how it behaves during a fall.
- An e-stop button is not automatically safety-rated. A wireless stop that depends on software and a network link may work well in practice, but it is a different thing from a redundant, analysed safety chain.
- Rules are changing. The EU Machinery Regulation, which applies from 2027, tightens expectations for machines with self-evolving behaviour; see our overview of the Machinery Regulation and humanoid robots.
This article is general information, not legal advice. Whether a particular robot meets the legal requirements for a particular use is a question for the manufacturer, the supplier and, where needed, a qualified adviser.
What this means for event organisers
- Ask what happens when the stop is pressed. Does the robot cut power, crouch or sit first, and how long does that take? A clear, demonstrated answer matters more than a label.
- Ask whether any stop function is safety-rated and to which standard. If the answer is "no rating", that can still be acceptable for a controlled demo, but plan distance and supervision around it.
- Ask what happens if the radio link drops. A robot that safely sits or stops when it loses its controller is easier to manage in venues with crowded Wi-Fi.
- Keep a clear zone around a standing humanoid. A robot that might fall needs space at least equal to its height around it. Put this into your public event safety plan.
- Confirm who holds the stop. A trained operator should keep the stop device within reach at all times; see who operates the robot and what staff you need.
Frequently asked questions
What does PL d or PL e mean on a robot?
It is a Performance Level under ISO 13849-1, graded from a (lowest) to e (highest). It applies to a single safety function, such as an emergency stop, and reflects its architecture, component reliability, fault detection and protection against common-cause failures. It tells you how reliably that one function works. It does not describe the safety of the whole robot or how it behaves in other situations.
Why can a humanoid robot not just switch off in an emergency?
A standing or walking humanoid keeps its balance by constantly adjusting its joints. If power is cut instantly, it can collapse and fall onto people or objects nearby. Researchers call this the fail-passive gap. Manufacturers therefore tend to bring the robot to a stop, settle or sit it down, and only then remove power, so that losing power no longer causes a fall.
Is a wireless e-stop on a rental robot safety-rated?
Not necessarily. Many robots use a software stop that depends on a radio link and a heartbeat signal. These can work well, and some robots stop automatically if the link drops, but they are not the same as a certified safety function with a documented Performance Level. Ask the supplier which stop functions, if any, have been rated and to which standard.
Is there a safety standard specifically for humanoid robots?
ISO is drafting ISO 25785-1, a safety standard for dynamically stable legged robots, including humanoids. Until it is published, manufacturers rely on general machinery standards such as ISO 13849-1 and IEC 60204-1, which were written mainly with fixed machines in mind. This is one reason why suppliers describe stop behaviour in different ways.
How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-11-15. Spotted an error? Write to info@humanoid4rent.com.
Quellen
Die Fakten in diesem Artikel wurden am Tag der Erstellung anhand dieser Seiten geprüft.
- IFA Report 2/2017e: Functional safety of machine controls – Application of EN ISO 13849 – DGUV
- SISTEMA Software zur Sicherheit von Maschinensteuerungen – IFA / DGUV
- Toward Certified Functional Safety for Industrial Humanoid Robots: The Fail-Passive Gap and a Feasibility Study – arXiv
- Learning Safe-Stoppability Monitors for Humanoid Robots – arXiv
- E-Stop Service – Spot 5.2.0 documentation – Boston Dynamics
- Digit 5 May Be the First Humanoid Robot Worker That's Truly Safe – IEEE Spectrum
Mehr aus dem Blog
Möchten Sie bei Ihrem nächsten Event einen Roboter?
Beschreiben Sie einmal Datum, Ort und die Aufgabe des Roboters. Anbieter antworten mit Angeboten, die Sie direkt miteinander vergleichen können.
Anfrage sendenSie arbeiten in der Robotik? Schreiben Sie für unseren Blog. →