Robotic Perception: One Box, Two Detections
A robotic perception model can detect two boxes where there is only one, with high confidence in both. A reflection is enough. Confidence is not reliability. Here is why, and what a calculated reliability for each prediction changes.
Why a Robot Sees Two Boxes Where There is Only One
Take a mobile robot in front of a real cardboard box and its reflection in a wall mirror. The perception system detects two boxes and gives each a high score. In this illustrative scene, the scores are 96% and 91%: these are not actual results. The robot might bypass a ghost object and lose efficiency. It could also follow an incorrect trajectory and hit the real box. A person would stop to look again. Nothing in the two scores prompts the robot to do the same.
Reflections and Shiny Surfaces
A reflection is a physical ghost: the second box is located elsewhere in space. Reflective surfaces are a documented source of error for both cameras and LiDARs. A camera can be confused by reflected content in a mirror (Yang et al., 2019). In experiments with a mobile robot and a 2D LiDAR, laser scans can be entirely reflected, objects can be mapped behind the mirror, and the robot can detect itself on its surface (Damodaran et al., 2023).
In a warehouse, a mirrored wall or a glass partition at the end of an aisle could produce the same effect. On the detection side, the second box is a reflection false positive, and the reported score does not say whether the detection is real. Vehicles encounter a similar problem: ghost walls in LiDAR point clouds. We cover perception errors in autonomous vehicles separately.
Occlusions and Duplicates
A duplicate is an artifact of the inference pipeline: multiple boxes overlap on the same object. Detectors produce numerous candidate boxes around each real object (Liu et al., 2019). Non-maximum suppression (NMS) keeps the highest-scoring box and discards any box that overlaps with it beyond a set threshold. Many detectors use this step, while some recent models do without (Ultralytics documentation, 2026).
A single threshold forces a trade-off. A low value discards real objects that overlap, creating false negatives. A high value lets more false positives through (Liu et al., 2019). This work focuses on pedestrian crowds: stacked cardboard boxes are therefore only an analogy. A threshold tuned for clear shelves might discard a real box in a stack, or keep two boxes around a single one.
The same object can also be output under two labels when suppression is applied class by class (Ultralytics documentation, 2026). A surviving duplicate carries a score just like any other box, and nothing in that score indicates that the object has already been counted.
Lighting Outside the Training Distribution
In a 2019 benchmark on corrupted images, standard object detectors retained only 30% to 60% of their original performance (Michaelis et al., 2019). The corruptions were synthetic, and the benchmark measured detection performance, not confidence. Furthermore, ReLU networks almost always produce highly confident predictions far from their training data (Hein et al., 2019). This study focuses on classification, not object detection.
The connection between these two facts is our inference, not a measured result. A change in lighting that moves images away from the training data could degrade detections without lowering the reported score. A loading dock lit differently than the aisle is a plausible scenario. See out-of-distribution detection and perception in degraded conditions.
Confidence is Not Reliability
A confidence score is what the detector says about its own output, and a softmax probability is the classic example. Reliability is what an independent layer measures about that output.
Confidence score. The confidence score refers to the number a detector itself attaches to each of its detections. In the original YOLO design, it is defined as the probability of an object being present, multiplied by the overlap between the predicted box and the ground truth. It is not a probability of being correct, measured in the field.
Reliability per prediction. Reliability per prediction refers to a measure, calculated for each individual prediction and before decision-making, of the confidence that can be placed in it. It is produced by a layer independent of the model. For a 2D detection, it takes the form of uncertainty about the position and size of the box (σx, σy, σw, σh).
In our view, a native score primarily indicates how much the input resembles what the model has learned. This is an interpretation, not a measured property of all detectors. On average, a high score remains associated with a correct detection.
Native confidence score | Reliability per prediction | |
|---|---|---|
What it measures | A learned output: object presence multiplied by box overlap (original YOLO design) | The uncertainty on the position and size of each box (σx, σy, σw, σh) |
Who produces it | The detector itself | An independent layer placed next to the model, without access to its internal workings |
When | During the forward pass of the model itself | At the time of prediction, before decision-making, in real time |
What the robot can do with it | Apply a global threshold: keep or reject | Proceed, slow down, observe again, or request verification, detection by detection |
What a Reliability Score per Frame Changes
Let's return to the scene. TrustalAI Vision quantifies the uncertainty of each detection in real-time. A reliability score per frame summarizes this uncertainty for each detection in an image, allowing a system to set thresholds. The real box carries low uncertainty: the robot proceeds. The reflection carries high uncertainty: the robot slows down, observes again, or adapts. TrustalAI does not tell the robot which box is a reflection. The robot's system decides, and this result is illustrative, not a guarantee.
Proceed, Slow Down, or Request Verification
Detection Reliability | What the downstream system can do |
|---|---|
High: low uncertainty on position and size | Proceed at normal speed |
Intermediate: higher uncertainty | Slow down, observe again from a different angle, or adapt trajectory |
Low: highest uncertainty | Flag the detection for human verification |
Robots already slow down and stop, but based on distance and zone, not on the reliability of each detection. Three examples show where a reliability signal would be plugged in:
Speed and separation monitoring is a functional safety function of collaborative applications, not AI monitoring.
The ISO 10218-2:2025 standard, dedicated to industrial robot systems and integration, describes it among its collaborative methods. It relies on measured distances, and these come from perception. Uncertainty about the detected position would tell the system how much to trust each distance.Safety laser scanners on mobile robots typically use a warning field that slows the vehicle down and a protection field that stops it. A reflection inside such a field could trigger the same reaction as a real obstacle: this is our inference. The reaction exists, but the reliability of each detection is missing.
A 2017 collision avoidance method proceeded cautiously in unfamiliar environments and accelerated where it had high confidence (Kahn et al., 2017). Its uncertainty relates to predicting a collision, not to object detection: it is a precedent for the principle, not a detection method.
Without a reliable signal, a robot is slowed down or placed under human supervision. With a reliable signal, it can maintain full speed where perception is reliable, with fewer unplanned stops, less downtime, and better operator safety. An unnecessary stop costs efficiency, and a false green light costs safety. See unplanned stops caused by AI vision and physical risk in shared zones.
The GPS Circle, Robot Edition
A GPS displays a blue dot and a circle around it. On the highway, the circle is tight. In a tunnel, it widens: the same system tells you it is less sure, instead of pretending otherwise. Now imagine a robot in a warehouse aisle. Behind a pallet or in front of a mirror, the blue dot looks the same, but the circle should be wide. TrustalAI draws this circle as a Reliability Bubble around each detection: the closer the bubble is to the box, the stronger the confidence in the object.
Most AI systems only ever show you the blue dot, even underground. TrustalAI adds the circle to any prediction, in real time.
Reliability per Prediction vs. Post-Mortem Monitoring
Existing tools work on aggregates. They wait for many predictions, calculate metrics, then report after the fact. This does not help a robot that must make a decision right now. The table puts the two approaches side by side.
Post-mortem monitoring | Reliability per prediction | |
|---|---|---|
Principle | Wait for several predictions, then calculate metrics | Calculate a reliability metric for each prediction |
Timing | After execution: the decision is already made | At the time of prediction, before decision-making |
Execution time | Not real-time | Real-time, less than 100 ms, 20 ms on edge |
Others analyze after execution: the decision has already been made, the action has already been taken. TrustalAI brings reliability at prediction time, before the decision is made.
For a robot, the difference is concrete. A dashboard that flags ghost detections after the shift tells you the robot has already veered. A global threshold is a setting. A reliability score is an answer for each detection. We compare both approaches in AI monitoring vs. reliability per prediction.
Results Measured with TrustalAI Vision
TrustalAI Vision adds the reliability layer to computer vision models.
80% fewer errors on position and orientation
92% fewer false positives
20 ms latency on CPU
8 ms latency on NVIDIA Jetson Orin
All results are measured in real operational conditions, against the baseline model of each product.
These are general product metrics, not figures specifically for reflections or duplicates. On edge computing, latency matters: see edge AI reliability under 20 ms.
Frequently Asked Questions
Why does a robotic vision model detect objects that do not exist?
Reflections can show a camera or LiDAR a second copy of a real object (Yang et al., 2019; Damodaran et al., 2023). A single overlap threshold can leave duplicate boxes around the same object (Liu et al., 2019). Images far from the training data can degrade detections (Michaelis et al., 2019). In all cases, the robot receives a score, not a proof.
Is a confidence score the same thing as reliability?
No. A confidence score is the number a detector attaches to its own output. In the original YOLO design, it combines the probability that an object is present and the box overlap (Redmon et al., 2016). Reliability per prediction comes from an independent layer, calculated for each prediction before decision-making.
Why does my detector return two boxes for a single object?
A detector typically proposes several candidate boxes per object, and non-maximum suppression discards those that overlap too much with the best box. A low threshold can discard real objects, while a high threshold keeps duplicates (Liu et al., 2019). The same object can also appear under two labels when suppression is done class by class (Ultralytics documentation, 2026).
Why does my robot stop in front of an obstacle that does not exist?
In experiments with a mobile robot and a 2D LiDAR, objects can be mapped behind a mirror's surface, and the robot can detect itself there (Damodaran et al., 2023). A planner that treats these returns as obstacles might stop or make a detour for nothing. This final step is our inference.
Can we add a reliability layer to an existing robotic perception model without retraining it?
Yes. TrustalAI is plug-and-play and works with black-box models. It reads your model's predictions in real-time and calculates their reliability, without retraining and without accessing your intellectual property. Measured latency is 8 ms on NVIDIA Jetson Orin and 20 ms on CPU (TrustalAI Client Deck V10, 2026). We do not extrapolate these figures to other hardware.
Share
Related articles
