Cobot Guarding Is Decided by the Task, Not the Model
Direct answer: A collaborative robot does not come with permission to remove fencing. The application decides: the end effector, the part, the speed and force, the layout, and how often people are inside the working area. Those inputs feed a task-based risk assessment, and the guarding decision is its output. An open station in a photograph is a layout, not a compliance conclusion.
Who this is for: This guide is written for automation engineers and safety coordinators specifying a collaborative application, and for managers who have to sign off the result.
Scope: It covers how the task, the tool, the part, and human exposure drive the guarding decision. It is not a risk assessment, and it does not replace the assessment that the integrator and the user of the complete application must produce.
Cobot guarding and risk assessment start with the application, not the model
Collaborative robots are built with features that make shared workspaces possible: power and force limiting, sensitive joints, rounded geometry, and controlled speed. Those features enable an application. They do not certify one. The safety case belongs to the complete installation, including the tool, the part, the fixture, the surroundings, and the way people interact with it.
In the reference footage for this guide, an arm with a fastening spindle works through a row of machined blocks on an open bench with only a profile frame around it. That shows how the station runs. It does not show, and must not be read as, evidence that the station is compliant, because the evidence for that lives in the assessment file, not in the picture. Cobot guarding and risk assessment therefore begin with the installed application and end with an acceptance record that names the residual risk.
The end effector usually carries the risk
Most of the hazard in a collaborative application arrives with the tool. A fastening spindle adds rotation and torque reaction. A gripper adds pinch points and a dropped-load scenario. A grinding or deburring head adds rotating abrasive, sparks, and dust. A dispensing head adds pressure and a substance. None of these are properties of the robot, and all of them change the answer about guarding.
Part geometry adds a second layer. Sharp edges, hot parts, long parts that extend the swept volume, and heavy parts that change the dropped-load consequence all belong in the assessment. In practice, the useful question is not “is this robot collaborative” but “what can this tool and this part do to a person who is standing here”. EVST specifies the end effector and its worst-case state before the arm, because that is where most of the hazard sits. According to ISO/TS 15066:2016, guidance for collaborative industrial robot systems addresses the collaborative workspace and the application within it, which is why the tool and the part are assessed alongside the arm.
Speed, force, and how often people are actually present
Reduced speed is a common measure, and it works, but it has a cost that has to be stated honestly: a slower application is a slower application. The engineering decision is where that cost is acceptable, and where a guard, a light curtain, or a scanner buys back the speed at lower total cost than permanent speed reduction.
Human exposure is the input teams most often estimate rather than measure. Count the real interactions per shift: loading, unloading, adjusting, clearing a fault, changing consumables, and inspecting. A station that people touch twice a shift and a station that people share continuously are different applications, even with the same robot and the same tool.
EVST specifies collaborative applications from that exposure count, because it changes the layout, the tooling, and often the choice between a collaborative arm and a small industrial arm behind a light guard. Similar reasoning appears in dispensing and path-controlled applications, where tool behaviour dominates the risk profile. According to ISO 12100:2010, risk evaluation follows hazard identification and risk estimation, which is why measured human exposure enters the application review before a protective measure is chosen.
Decision table: application evidence to guarding approach
The table is an entry point to the assessment, not a substitute for it. The obligation to assess remains with the integrator and the user of the complete application.
| Application evidence | Guarding approach it points to | What the assessment must still show |
|---|---|---|
| Light part, blunt tool, rare human presence | Open station with speed and force limits | Contact scenarios for every reachable body region and the tool in its worst state |
| Rotating or abrasive tool | Local guarding of the tool, or a guarded cell | Ejection, entanglement, spark and dust control, and maintenance access |
| Heavy or sharp part | Guarding plus fixture retention | Dropped-load path, clamping failure mode, and swept volume with the part |
| Continuous shared workspace | Presence sensing with zoned speed | Detection coverage, stopping performance, and restart control |
| Task or tooling not yet fixed | Hold the decision | Freeze the task, tool, part, and layout before assessing |
The documents that frame the decision
ISO 12100:2010 provides the general principles for design, covering hazard identification, risk estimation, risk evaluation, and risk reduction. It is the backbone of the assessment and the reason the analysis starts from tasks and hazards rather than from equipment categories.
For robot applications and cells, ISO 10218-2:2025 addresses integration, commissioning, operation, maintenance, and decommissioning requirements. For collaborative operation specifically, ISO/TS 15066:2016 provides guidance for collaborative industrial robot systems and the collaborative workspace, including the considerations that apply when contact between a person and the application is foreseeable.
Naming those documents in the specification does not make an application safe. What makes it defensible is the assessment record that maps each task to its hazards, its measures, and the residual risk that the user accepts and trains for. According to ISO 10218-2:2025, integration, commissioning, operation, maintenance and decommissioning requirements apply to robot applications and cells, which is why acceptance of the station is documented rather than inferred from the robot category.
Writing the assessment into the specification
A specification that only names a robot model leaves the most important part of the application undefined. A useful specification states the task, the end effector and its worst-case state, the part, the layout with human positions, the expected interaction frequency, and who produces and maintains the assessment file. Those items are what a supplier needs in order to quote something real.
It also helps to state what happens when the application changes. Tooling gets swapped, parts get revised, speeds get raised because a target moved, and layouts get rearranged around new equipment. Naming, in advance, which of those changes triggers a re-assessment prevents the slow drift where an approved application quietly becomes a different one.
Training and the residual risk statement belong in the same document. An application that is sound as designed, but operated by people who were never told which measures they depend on, is not the same application. The assessment file should end with what the operator has to know, not with a conclusion nobody reads.
Patterns that clear review and patterns that do not
Applications that clear review tend to share a shape: a blunt or enclosed tool, a light part with a defined retention method, a small and well-understood swept volume, and human interaction that is scheduled rather than continuous. In that shape, speed and force measures are doing work that matches the actual hazard.
Applications that do not clear review tend to share a different shape: a tool that carries its own hazard regardless of arm speed, a part whose loss or contact would injure independently of the arm, or a layout where people pass through the working area for reasons unrelated to the station. In those cases the answer is usually local guarding, presence sensing, or a rearrangement that removes the traffic, not a further speed reduction.
The practical lesson is that the robot category is the least informative fact in the file. What decides the outcome is the tool, the part, the layout, and how often a person is inside the space while the application runs.
Layout choices that change the answer
Guarding is often treated as something added at the end, but layout decides how much guarding the application needs in the first place. A station where material arrives from behind the robot, and where the operator loads from a position that is never inside the swept volume, needs less protective equipment than the same task arranged so that people reach across the working area.
Three layout patterns tend to reduce risk without slowing the application. Separating the load position from the process position lets a person present work while the arm is elsewhere. Arranging the swept volume so it faces a wall or a fixed structure removes a whole set of approach directions. Keeping consumables, tooling, and waste outside the working area removes the routine reasons a person would step into it.
The opposite patterns are equally recognisable: a bench where the part, the tools, and the finished goods all sit inside the arm’s reach; a station placed in a walkway; or a fixture that requires an operator to steady a part during the cycle. Each of those converts an occasional interaction into a continuous one, and each pushes the assessment toward heavier protective measures.
This is why layout should be reviewed together with the tool and the task rather than after them. Changing a bench arrangement is cheap while the station is on paper and expensive once it is installed and producing. EVST reviews layout together with the tool and the task, since a bench arrangement is cheap to change on paper.
Frequently asked questions
Can a collaborative robot always run without a fence?
No. The absence of a fence is a conclusion of a completed risk assessment for the whole application. Change the tool, the part, the speed, or the layout and the conclusion has to be revisited.
Does a slower cobot always make the station acceptable?
Reduced speed lowers some risks, but it does not address entanglement, ejection, sharp edges, hot parts, or dropped loads. Those need their own measures regardless of speed.
Who is responsible for the assessment?
Responsibility sits with the party that integrates and the party that uses the complete application, not with the robot supplier alone. Contracts should state who produces the assessment file and who maintains it after changes.
What triggers a re-assessment?
Any change to the tool, the part, the fixture, the layout, the speed regime, or the way people interact with the station. Program changes that alter the swept volume also count.
Project inputs for an application review
Send the following and the guarding approach and assessment scope can be reviewed for the actual application:
- the task description, cycle by cycle, including manual interventions
- the end effector and its worst-case state, including tool changes
- part mass, geometry, edges, temperature, and how it is retained
- the layout with human positions and how often people enter the working area
- the target cycle and who will own the assessment file after handover
Send the task description, the end effector and its worst-case state, the part mass and geometry, the layout with human positions and access frequency, and the cycle you are targeting. That set is enough to review the guarding approach and the assessment scope for the application. Related reading: workstation configuration reference, wider robotic factory integration, robot product range.