background gradient shape

AI

Robot cadence: perception reliability sets your line speed

Robot cycle time: perception reliability sets your line speed

Robot throughput depends on how much trust the cell places in what it sees. Every doubtful detection is paid for with a protective stop, reduced speed, or human intervention. Measuring the reliability of each frame allows nominal speed to be maintained wherever perception is trustworthy.

Why your robots run below their nominal speed

A mobile robot travels up a warehouse aisle. Sunlight enters through a dock door and casts a shadow across the floor. The robot stops dead in its tracks. There is nothing in front of it, and an operator has to come restart it. In the neighboring cell, an arm has been running below its nominal speed ever since commissioning, set that way just in case.

The scene is illustrative, but commonplace. None of these robots are broken. They are being cautious because they do not know which detections to trust, and this caution takes four forms.

Protective stops triggered for nothing

A phantom obstacle causes no damage, but it stops the robot. For autonomous mobile robots (AMRs), sensor manufacturer ifm cites dust, highly reflective surfaces, and sunlight among the causes of false detections (ifm, accessed October 21, 2026). Floors range from concrete to glossy tiles. The same manufacturer sums up the problem with one question: how do you explain to your customer that a robot stopped for no identifiable reason?

Raising a global detection threshold eliminates unnecessary stops, but takes legitimate stops with it. A single setting cannot tell a shadow from a cardboard box, as we demonstrated with a robot seeing two boxes instead of one.

Speed capped as a precaution

Without knowing which detections to trust, the system cannot modulate its speed on a case-by-case basis. The system integrator therefore sets up the cell at commissioning based on the worst-case scenario they can anticipate, and that setting remains. Reduced speed becomes the standard operating rate. It takes its toll on every cycle, even when perception was perfectly reliable.

Nothing flags this cost. The displayed cycle time looks normal, and no one compares it to nominal speed anymore.

The degraded sensor that no one sees

A dirty lens does not trigger an alarm. The model continues to produce detections, simply less accurate ones, and the cell compensates with more stops and more human checks.

The automotive industry has taken the problem seriously. The public WoodScape dataset counts camera soiling detection among its nine tasks (Yogamani et al., 2019). The authors of SoilingNet note that cameras mounted on the vehicle exterior get dirty, and that their performance degrades more than that of other sensors. We cite this automotive work as an analogy. A pallet never seen before has the same silent effect: seeing out-of-distribution detection.

Human supervision that never goes away

The ultimate safety net is human. Someone watches a screen all day and restarts the robot after every questionable stop. This role was supposed to last only for the pilot phase. As long as no one knows which detections to trust, it stays, and the pilot never transitions into full production. AI vision-related line stops share the very same origin.

The reliability tax: where robot throughput is lost

The reliability tax refers to the throughput that a robotic cell gives up as a precaution because it does not know which detections to trust. We call it a tax because it is levied on every cycle without anyone ever explicitly choosing to pay it. It appears on no invoice. Yet it can be found in two places: the collaborative robot separation equation and the performance category of OEE.

What monitoring tells us

Monitoring refers to a collaborative method described by the technical specification ISO/TS 15066: the robot constantly maintains a protective separation distance from the operator. According to the presentation given by Svarny et al. (2019), this distance adds up six terms:

  • the operator's movement during the robot's reaction and stopping time, at a default walking speed of 1.6 m/s;

  • the distance traveled by the robot during its reaction time;

  • its stopping distance;

  • the distance a body part can move into the detection zone before being detected;

  • the operator's position uncertainty, stemming from the measurement tolerance of the sensing device;

  • the robot's position uncertainty, stemming from its position measurement accuracy.

Measurement uncertainty is therefore added to the required distance. It cannot be compensated for: it is paid for in safety margins. Let's look at the formula. For a given available distance, every margin added to cover an uncertainty leaves less room for the terms that depend on robot speed. This is an interpretation of the equation, without empirical measurements attached.

This distance can be calculated continuously or fixed once and for all based on the worst-case scenario (Svarny et al., 2019). A worst-case value is paid for at every cycle. In practice, there are two thresholds: below the first, the robot slows down; below the second, it stops, then resumes once the path is clear. This is the intended safety response. The cost arises when it triggers for nothing, or too early.

The same authors timed the difference on a simulated pick-and-place task, using an RGB-D sensor to track the operator. The task took 154 s at full speed without an operator, 256 s at permanent reduced speed, and 267 s with zone-based stopping around the base. Thresholds based on the distance between key points on the operator and the robot brought it down to 231 s and 228 s (Svarny et al., 2019). This is a third-party laboratory experiment on a single cell and a simulated task, and cannot be generalized. It merely shows that, in this setup, a finer estimation of operator position reclaimed cycle time.

This equation falls under the certified safety chain, which we do not alter. The same logic applies to the operational layer: planning, navigation, and decisions to slow down made before the safety function needs to intervene. Here too, a system that does not know which detection to trust slows down everywhere.

Where it appears in your OEE

OEE (Overall Equipment Effectiveness) is an indicator combining three factors: availability, performance, and quality (OEE.com, accessed October 21, 2026). Among the six big losses defined in TPM, two relate to the reliability tax:

  • reduced speed: the equipment runs slower than its ideal cycle time;

  • small stops: short pauses, typically one or two minutes, cleared by an operator.

Both are performance losses. They are easy to overlook because the machine is not broken down: it is simply slower. This is the section of the OEE where the reliability tax can be seen.

A per-frame reliability score to govern speed

TrustalAI Vision computes a reliability metric for every detection, in real time. So the robot slows down only when it should. The per-frame reliability score turns this into a speed policy, and thus a robot throughput policy. Instead of a single operating mode set for the worst-case scenario, the cell adapts its pace to what it sees.

What the reliability layer sees

What the cell does

Effect on throughput

High reliability on the frame

Nominal speed

No slowdown during these cycles

Uncertainty across a few frames

Targeted slowdown, limited to those frames

Speed resumes as soon as reliability recovers

Persistent uncertainty

The robot hands control over to a human

The operator only intervenes in cases that require it

Faulty sensor

Flagged immediately

The sensor issue is addressed instead of being compensated for by lower speed

This table outlines a control logic without quantified gains: each integrator sets their own thresholds based on their specific data.

The reliability layer does not replace certified safety functions. Laser scanners, protective stops, and speed and separation monitoring remain in place, establishing the safety floor on their own. Nominal speed remains what the safety configuration permits. TrustalAI operates on top, within the operational layer. The safety aspect is covered in our article on physical risk in shared workspaces.

In congested aisles, TrustalAI V-Tracking complements Vision: real-time, deterministic, learning-free multi-object tracking that keeps each person and each cart on a single trajectory. To evaluate this tracking without annotations, see reliability without ground truth.

Per-prediction reliability vs. post-hoc monitoring

Existing monitoring tools work on aggregated data. They wait for numerous predictions, compute metrics, and then report after the fact. When it comes to throughput, that is too late: an end-of-shift report will not bring back lost speed.


Post-mortem monitoring

Per-prediction reliability

Principle

Wait for multiple predictions, then compute metrics

Compute a reliability metric for each prediction

Timing

After execution: the decision has already been made

At the moment of prediction, before the decision

Execution time

Not real-time

Real-time, under 100 ms, 20 ms at the edge

Others analyze after execution, once the decision has already been made and action taken; TrustalAI provides reliability at the moment of prediction, before the decision.

A dashboard may reveal the next day that a robot frequently stopped for no reason. It does not tell the robot which detection to trust at the moment it selects its speed. Per-prediction reliability answers that question, frame by frame. See AI monitoring vs. per-prediction reliability and per-prediction reliability versus global performance. Model drift follows the same rule: detected late, it has already cost you.

Results measured with TrustalAI Vision

TrustalAI Vision adds a reliability layer to existing computer vision models

  • 92% fewer false positives

  • 80% fewer position and orientation errors

  • 8 ms latency on NVIDIA Jetson Orin

  • 20 ms latency on CPU

All results measured in real operating conditions against each product's baseline model.

For a robot, we interpret these as follows. Fewer false positives mean fewer phantom obstacles and fewer unplanned stops. Fewer position and orientation errors yield safer movements around people and equipment. At 8 ms, the layer runs on the robot's existing onboard compute.

These are general product metrics, not measured throughput gains. They support running the robot at full speed safely, as well as operator safety.

No retraining, runs on the robot's existing compute

TrustalAI is plug-and-play. It connects alongside your perception model—which remains a black box—and reads each prediction in real time without accessing weights or training data (TrustalAI Client Deck V10, 2026). No retraining, and no access to your intellectual property. The layer runs at the edge on the robot's onboard computer, or in the cloud.

  1. Your model predicts.

  2. TrustalAI evaluates the reliability of each detection.

  3. Your system acts: it maintains speed on reliable detections, and slows down or raises alerts on the rest.

Full confidence on reliable detections. Caution only on the rest.

TrustalAI adds on as a reliability building block: the cell keeps its sensors, its model, and its safety configuration.

Frequently asked questions

Why isn't my robot running at its nominal speed?

One of the reasons comes down to perception: the system does not know which detections to trust. The integrator sets the speed for the worst-case scenario, and the robot stops or slows down whenever a detection seems questionable. A dirty sensor or ongoing human supervision amplifies this loss. We call this cost the reliability tax. It appears in the performance category of your OEE.

How is the protective separation distance of a collaborative robot calculated?

According to ISO/TS 15066, as presented by Svarny et al. (2019), it first adds the operator's travel during the robot's reaction and stopping time, the robot's reaction distance, and its stopping distance. Added to this are an intrusion distance and two position uncertainties: the operator's, tied to the sensor, and the robot's. Any added uncertainty increases the required margin.

Should detection sensitivity be lowered to prevent nuisance stops?

Not without trade-offs. Raising an overall detection threshold eliminates unnecessary stops, but also legitimate ones: a single setting cannot tell a shadow from a cardboard box. False detections in AMRs have various causes, such as dust, highly reflective surfaces, or sunlight (ifm, accessed October 2026). A per-frame reliability score handles each detection based on its reliability, without touching the safety configuration.

Is reduced robot speed counted in OEE?

Yes. Reduced speed and small stops are part of the TPM six big losses, representing two types of performance loss (OEE.com, accessed October 2026). A robot running slower than its ideal cycle time is not broken down: its availability remains intact, but its performance drops. That is why this loss often goes unnoticed during production reviews.

Does a reliability layer replace robot safety functions?

No. Certified safety functions—such as area scanners, protective stops, and speed and separation monitoring—remain in place and independently determine the safety baseline. A reliability layer like TrustalAI operates on the operational layer: planning, navigation, and decisions to slow down. It helps the cell slow down only for questionable detections, without altering the safety chain.

Share

Gradient Circle Image
Gradient Circle Image
Gradient Circle Image

Make your AI reliable now

Make your AI reliable now

Make your AI reliable now