Topics Safety · Dossier
A humanoid robot shares space with people. Seven safety layers work together to make that safe. Turn the robot, tap the points and send a person closer.
Example values for illustration
Seven safety layers
No single measure is enough. Safety comes from layers that back each other up: if one fails, the next one takes over.
01 / 07
Protective fields around the robot
Virtual fields surround the robot. When someone enters a field, the robot adapts its speed, all the way to a stop. The technical term is speed and separation monitoring.
Define walkways, work areas and zones with public traffic. They determine how large the fields must be.
02 / 07
Sensors detect people all around
Cameras, depth sensors and LiDAR capture the surroundings. Only sensors with proven reliability count for protective functions. AI image recognition helps on top but does not replace them.
Check lighting, reflections, dust and tight corners on site. What the sensor cannot see must be secured physically.
03 / 07
Light, gaze and voice announce
People feel safe when they know what the robot will do next. A status light, eyes in the visor and short voice cues make its intent readable.
Agree on one signal language for all sites and train it like traffic signs.
04 / 07
Arms and hands stay gentle
If contact does happen, it must not injure. Joints measure their torque, soft covers absorb impact, and limits for force and pressure depend on the body region.
Assess every gripping task on its own. A cardboard box is harmless, a knife or a hot cup is not.
05 / 07
Keep balance, fall in a controlled way
A two-legged robot cannot simply cut power, or it will topple. It needs a safe stop while standing and, if it falls anyway, a controlled fall away from people.
Check floors, ramps, cables and steps. Slippery or uneven surfaces belong in the risk assessment.
06 / 07
Emergency stop always within reach
Anyone can stop the robot at any time: with the emergency stop button on its body, a wireless handheld or through the installation. A dedicated safety controller monitors this, independent of the AI.
Make it clear: everyone may stop the robot, only trained people may release it again.
07 / 07
AI with guardrails, data protected
The AI plans tasks, while a separate safety layer sets hard limits for speed, force and zones. Add protection against attacks, privacy for camera images and complete logs.
Involve IT security, data protection and employee representatives early. Decide which data is stored and for how long.
Transfer
The company is the starting point. The same principles apply in care, retail, service and learning, only weighted differently.
The core case: robot and team work in the same hall.
Closeness is intended, so force and speed count double.
Many people who have never had a briefing.
Friendly in manner, safe at all times.
The robot as subject and companion of learning.
The same seven layers, weighted differently. Three dots mean especially critical.
| Layer | Production & logistics | Care | Retail | Service | Learning |
|---|---|---|---|---|---|
| 1Keep distance | especially critical | very important | especially critical | very important | very important |
| 2Perceive | especially critical | very important | especially critical | very important | very important |
| 3Show intent | very important | especially critical | very important | especially critical | especially critical |
| 4Limit force | very important | especially critical | very important | especially critical | especially critical |
| 5Stay stable | especially critical | especially critical | very important | very important | very important |
| 6Stop safely | especially critical | very important | very important | very important | especially critical |
| 7Decide safely | very important | especially critical | especially critical | very important | especially critical |
Focus on companies
How to introduce humanoid robots in your organisation. The steps apply just as well to a care home, a store, a hotel or a school.
Which task, which place, which people? A clearly defined case is easier to secure than an all-rounder.
Record hazards systematically: motion, force, falls, loads, data. The result determines the protective measures.
Select standards, calculate protective fields, validate safety functions. Integrating the robot into an installation often creates a new machine with its own CE marking.
Bring in employee representatives, data protection and IT security early. Openness about cameras and data builds trust in the team.
Start in a marked area with a few trained people. Measure, observe, adjust.
Who operates, who releases, who maintains? A short briefing for everyone who meets the robot.
Report incidents and near misses, review updates, renew the risk assessment after every change.
Rules
As of October 2026. There is no complete standard just for humanoid robots yet, so existing rules are combined.
Safety of industrial robots and robot systems, including collaborative applications.
Biomechanical limits for force and pressure when touching people. Its content has been incorporated into ISO 10218-2:2025.
First standard for dynamically stable industrial mobile robots, such as two-legged humanoids.
Safety of personal care robots. An important reference for care and service.
Functional safety of control systems, rated as performance level or SIL.
Replaces the Machinery Directive and for the first time covers machines with self-learning behaviour.
Obligations for AI systems. AI used as a safety component of a machine counts as high-risk.
Cybersecurity of connected products and protection of personal data such as camera images.
This overview is not legal advice and does not replace a risk assessment by experts.
FAQ
They move mass and force like any machine. Protective fields, force limiting, an emergency stop and a careful risk assessment reduce the risk to an acceptable level. What counts is the specific use.
Not a complete one yet. ISO 25785-1 is being developed for dynamically stable industrial mobile robots. Until then, ISO 10218, ISO 13482 and functional safety standards are combined.
No. Emergency stop, speed and force limits run in a separate, verified safety controller. The AI plans within these limits but cannot move them.
It first tries to keep its balance. If that fails, it falls in a controlled way and away from people where possible. That is why floors, loads and routes belong in the risk assessment.
With a clearly defined use case, a risk assessment and a pilot in a defined area. Employee representatives, data protection and IT security should be involved from the start.