What Actually Makes a Collaborative Robot Safe?

Updated June 2026 · Reviewed by the CCH Sensing technical team.

Collaborative robot safety is the set of design limits, collaborative operating modes, and risk controls that let a robot share a workspace with people without injuring them. It’s also one of the most misunderstood ideas in factory automation. Marketing says a cobot is “inherently safe,” so you can take it out of the box, stand next to it, and skip the fences. In reality, safety isn’t a property the robot carries around with it, it’s a property of the application: the arm, the tool on the end of it, the part it’s moving, the speed it run, and the people who share its space. Get that distinction wrong and a “safe” cobot will still send someone to the emergency room.

This guide is written by a safety-sensor manufacturer, so it is deliberately blunt about where power-and-force limiting stops and where an external light curtain, area scanner, or safety relay has to take over. We will cover the four collaborative operation modes from ISO/TS 15066, the body-region force limits that decide whether contact is acceptable, the standards stack (including the major 2025 revision of ISO 10218), how to run a risk assessment, and the five dangerous assumptions that get people hurt.

What Collaborative Robot Safety Really Means

A collaborative robot, cobot for short, is a robot designed to work alongside humans without a traditional safety fence. But “designed to” isn’t the same as “is.” Here’s the principle that everything else in this guide hangs on, what we call The Application-Safety Principle: a cobot is never safe or unsafe on its own; the application is what gets risk-assessed and signed off. That same arm is safe carrying a foam probe and dangerous carrying a deburring spindle.

That matters because the robot is only one piece. Under the standards, the bare robot is “partly completed machinery”; the thing that has to be made safe is the complete cell, robot, end effector, workpiece, layout, and the people. The U.S. Occupational Safety and Health Administration is explicit that there’s no robot-specific OSHA standardrobots are regulated under general machine-guarding rules (29 CFR 1910.212) and the General Duty Clause, and “a combination of controls and safeguarding should be used.”

What makes a cobot safe?

Four things, in order: (1) the robot’s built-in safety functions, power, force, torque and speed monitoring that trigger a protective stop; (2) the collaborative operation mode you choose for the task; (3) any external safeguarding the risk assessment calls for; and (4) the risk assessment itself, which ties the first three to your tool, payload and layout.

Skip the fourth and the first three are just features on a spec sheet. Those built-in safety functions, the safety features and protective features of the robot system, the extra safety measure an integrator adds, and the safety mechanisms that fire on a collision all combine so the machine can run around people, yet none of them removes the duty to assess how human workers actually come into contact with the robot. Field experience backs this up: lighter cobots stop on light contact, but “larger ones with more inertia” hit harder, so mass and payload change the answer.

The Four Types of Collaborative Operation (ISO/TS 15066)

ISO/TS 15066the technical specification that supplements ISO 10218, defines four collaborative operation modes. An application can combine them, and each one implies a different stop trigger and a different safety device. This is the table the top search results describe in prose but never lay out:

Universal Robots, which helped write the standard, notes that in practice only SSM and PFL deliver genuinely fence-free collaborative applications; the monitored stop is really a faster-restarting guarded cell, and hand guiding, driving the arm through hand-guided controls, is a teaching aid. Across collaborative robot applications, human-robot contact is the variable that decides the mode, so pick the mode for the task, not the other way round.

Power & Force Limiting: The Body-Region Force Ceiling

Power and force limiting is the mode people mean when they say “cobot.” It works by guaranteeing that any contact stays below the point where a human feels pain or injury. Those thresholds are not a marketing number, they come from ISO/TS 15066 Annex A, which sets maximum permissible force and pressure for 29 specific body locations (grouped into 12 body areas), based on pain-onset research. We call this reference The Body-Region Force Ceilingand almost no competing guide prints it:

Two numbers in that table do the most work. First, the face at 65 N is the lowest force ceiling on the body, which is why a cobot working near head height needs extra scrutiny. Second, the standard splits every limit into transient (a free impact, where the body can move away) and quasi-static (clamping or trapping, where it can’t). Quasi-static ceilings run roughly 40–65% of the transient values, because sustained compression damages tissue at lower force.

Practically, PFL cobots run between 250 and 1,000 mm/s, and the permissible speed falls as the effective mass (robot + payload + gripper) rises. Engineers patenting this field, for example US 11,453,122 B2, a torque-sensor-per-axis collaborative system, are all chasing the same thing: detect the contact fast enough to stop under the limit. In practice the amount of force and the torque applied at every joint must stay under the allowable ceiling, and the end-of-arm tooling, an electric or pneumatic gripper, counts toward it, because a sharp tool raises the risk of injury even when the arm itself is compliant.

Speed & Separation Monitoring: Letting a Cobot Run Faster

The trouble with PFL is that staying under the force limits forces the robot to crawl, Universal Robots admits the speed “will likely be too low to be useful for high-risk applications.” Speed and separation monitoring fixes this by keeping a measured gap between human and robot. When the operator is far away the robot runs at production speed; as they approach it decelerates; if the gap drop below the minimum protective separation distance, it stops.

That distance is calculated, not guessed. ISO 13855 gives the form S = K·T + C, where K is the human approach speed (the standard value is 2,000 mm/s), T is the total system reaction time (sensor response + robot stopping), and C is an intrusion/uncertainty allowance.

SSM is where an external sensing layer earns its keep. The NIST testbed work and the peer-reviewed SSM literature both stress that the sensing must itself be safety-rated and validated, not a standard automation scanner repurposed for safety. Our safety laser scanner guide walks through zone configuration in detail.

The Standards Stack: ISO 10218, ISO/TS 15066 and ANSI/RIA R15.06

Three documents govern cobot safety, and they’re layered, not interchangeable:

The hierarchy underneath them matters too: risk assessment follows ISO 12100, and safety functions must hit a performance level (PL d per ISO 13849-1) or safety integrity level (SIL 2 per IEC 62061) appropriate to the risk. Roberta Nelson Shea, Universal Robots’ Global Technical Compliance Officer, long-time convenor of the ISO robot-safety committee, and the leading authority on these documents, has spent decades stressing that the standard is a floor, not a ceiling.

When a “Fenceless” Cobot Still Needs Guarding

Here’s the part the cobot brochures skip. A power-and-force-limited robot manages blunt contact, it does nothing about a sharp, hot, or heavy end effector. Robotiq’s own palletizing example makes the point: moving multi-kilogram boxes at head height is “intrinsically unsafe,” so above a crawl that cell require a safety device regardless of PFL.

Are cobots safe to work without a fence?

Sometimes, and only when the risk assessment proves it for that exact tool, payload, speed, and layout. Trigger questions: can the end effector cut, burn, or crush? Can a body part be trapped against a fixture? Does the cycle need to run faster than PFL allows?

If any answer is yes, add a presence-sensing layer, a safety light curtain at a load station, an area scanner for floor zones, and a safety relay to execute the monitored stop. Each device must carry a rating matched to the risk (PL d/SIL 2, or Type 4 PL e for finger protection). For a side-by-side of the presence devices, see our scanner vs. light curtain comparison and the broader machine-guarding light curtain overview.

The 6-Step Cobot Risk Assessment Walk

Every standard points back to the same task-based process from ISO 12100. Here is The 6-Step Cobot Risk Assessment Walk we use, condensed:

Done well, the process is really risk reduction in stages: you reduce the risk by choosing the mode first, then layer safety controls and a safety system sized to the specific safety requirements and safety needs of that robotic system. Operators still need a robot safety training program covering the cell’s safety protocols before the line run. One discipline most teams miss: the risk assessment is a living document, so a new gripper or a new part program changes the risk profile and forces a review, a point industry trainers and groups like SACA keep repeating.

5 Fenceless-Cobot Assumptions That Get People Hurt

From field experience and the standards, these are the assumptions worth auditing before anyone stands next to the arm:

• ⚠️“A cobot is safe out of the box.” The robot is partly completed machinery; only the assessed cell is safe.
• ⚠️“PFL means no risk assessment.” Force limiting addresses blunt impact only, it says nothing about your tool, and the assessment is still mandatory.
• ⚠️“Slow always means safe.” A slow arm with a blade, a hot tip, or a clamping fixture still injures.
• ⚠️“Any gripper and payload are fine.” Effective mass sets the speed ceiling; a sharp or heavy payload can void PFL safety entirely.
• ⚠️“A CE-marked robot means a compliant cell.” The robot’s mark covers ISO 10218-1; the integrated cell needs its own conformity under 10218-2.

Why isn’t cobot safety as simple as it first appears?

Because the marketing answers a different question than the standard does. “Is the robot safe?” gets a confident yes; “is this robot application safe?” gets “it depends on the tool, the speed, and the people.” Applied Manufacturing Technologies puts the most common error plainly: the misconception that operators can share the same space as a PFL robot with no guarding “isn’t always” correct.

A peer-reviewed socio-technical review reaches the same conclusion, machines don’t always behave as assumed, so the inherent-safety belief has to be tested, not trusted.

Cobot Safety by Application: The Four-Mode Fit Test

Different jobs land on different modes and different devices. Use The Four-Mode Fit Testmatch the application to its likely mode and the safeguarding it usually needs:

Whatever the application, the safe-stop function in the background is usually a safety relay module tying the sensors to the robot’s safe inputs. For a full traditional robot-cell layout, the caged cousin of the collaborative cell, see our safety for robot cells solution.

“The question we ask integrators is never ‘is the cobot safe?’ — it is ‘what does the end effector do, and how fast?’ On a finger-protection load station we still specify a Type 4 PL e light curtain even on a force-limited arm, because PFL does nothing about a 200 N/cm² pinch from a sharp gripper.”

CCH Sensing Application Engineering team

Industry Outlook: What ISO 10218:2025 Changes for Your Next Cobot

The most important recent change is regulatory, not technological. In January 2025 the third editions of ISO 10218-1 and ISO 10218-2 were jointly published, and they make functional-safety requirements explicit rather than implied, and fold the collaborative content that lived in ISO/TS 15066 into the 10218 series itself. Shortly after, ANSI and A3 published a revised R15.06 adapted directly from the 2025 ISO editions.

For a buyer or integrator, the action item is concrete: stop treating ISO/TS 15066 as a standalone spec, and confirm which edition your robot OEM and your integrator are certifying to on your next install. A cell designed and documented against the 2011 standard is not automatically wrong, but during a transition window auditors and customers will increasingly expect the 2025 baseline. For context on scale, the reason the standards bodies moved, the IFR’s World Robotics 2025 report counted 542,000 industrial robots installed in 2024, the fourth straight year above 500,000; that volume of human-adjacent automation is what drove the rewrite (market figures here are background only).

On the horizon, skin-like tactile sensors and AI-vision human tracking promise faster, finer detection and better safety performance, but neither is yet certified for advanced safety use, so across real use cases the compliant fence-free cell still rests on the four modes, the force ceilings, and a documented risk assessment. Robotic safety of industrial robots and robot systems, collaborative or not, is still earned application by application.

Frequently Asked Questions