Sample
Sheet EA-03D · Solid Register

Sensor Deception Library

Some environments make a sensor look at one thing and see another, and the robot gets it wrong. Eight typical classes are collected here: standing water, specular surfaces, dark surfaces, negative obstacles, grating, steam, light transitions and repeating arrays. Each one is shown three ways: what a person sees, what the sensor receives, and what is actually there.

SheetEA-03
EntriesDS-01 – DS-08
/ 8 causes
Live solveDS-04, 1 entry
/ spatial only
Data basisMechanism real
/ figures indicative
Drawing no.EA-ATLAS-PV-2026-S
HOW TO READ

Three Parallel Views of One Scene

Person, sensor, fact. All three views are the same site from the same spot. Only the observer changes. Channel A is what a person sees with their eyes, Channel B is what the sensor actually receives, Channel C is what is actually there. Most misjudgements happen between B and C. The robot acts on B, and B is not always C.

On the right the three drawings are separated in depth in a view you can turn. Drag it, or press Home to face it square on, and the three ground lines snap into one. That is the moment it becomes obvious they were always the same site, just recorded by different observers.

Two buttons

Stacked separates the three drawings in depth. Drag to turn, or use the arrow keys, and Home squares them up. Side by side lines them up, which is better for close reading and for printing. Narrow screens always get side by side.

Verdict bands, effect by platform and how to build the red-team scene come with each entry below.

Data basis · measured against inferred

Physical cause, which way the misjudgement runs, and how it varies by platform are qualitative conclusions and can be read as they stand. Every specific figure for distance, thickness, reflectance and concentration is indicative, not measured.

The live solve in DS-04 computes spatial position only. Where a beam lands is worked out from mounting height, distance to the trench and angular resolution, and that part is real. It does not include material reflectance, signal strength or a real filter threshold. It shows you whether the mechanism happens, not whether your robot will fall in. That second question is what the L4 perception validation report answers.

DS-01 · SPECULAR / STANDING WATER

A Patch of Standing Water after Rain

On the service roads of a solar-plus-storage site, rain always leaves a few low spots holding water. A person walks around it. A robot may slam to a stop.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONdark, glossy groundservice road · after rain
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNdepth camera 0.9 m≈0.7 mmirror surface, no return → read as a deep pit
CHANNEL CGROUND TRUTH · what is actually there
SECTION30 mmwater filmC30 concrete road · bearing capacity normal
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person sees a patch of dark, glossy ground that is plainly flat.
Channel B · return · what the sensor receivesA mirror sends nearly all the energy off in one direction, so the sensor gets nothing back. The few rays that do return come in obliquely, get resolved along the reflected path, and land below the ground plane.
Channel C · ground truth · what is actually thereThe water film is 30 mm deep, and the concrete beneath it is intact.
Ground truth
Passable 30 mm of water, road bearing capacity normal
Perception verdict
Not passable no valid return over the water, and after depth completion it is read as a pit around 0.7 m deep
Mismatch
False-positive obstacle walkable ground called impassable
Consequence
Detours and repeated replanning. At worst the robot stops and starts along the edge of the water until the task times out.
Trigger conditions
0–6 hours after rain · low spots and settlement · most pronounced below 30° incidence.
WR-220 wheeled inspection moderateQR-X4 quadruped moderateUAV-M6 UAV n/aCL-P2 rail cleaning n/a
Red team · how this one is built

Wet the same stretch of road to a 20–40 mm film and run it three times, at morning, midday and evening sun angles. Record the moment the verdict flips.

DS-02 · SPECULAR / GLASS & POLISHED METAL

Glass Curtain Wall and Polished Cabinet Doors

The curtain wall on a control building. The stainless steel door panels on an inverter cabinet. A person knows at a glance it is a wall. To a lidar there is almost nothing there.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONcontrol building curtain wallinspection robotreflection
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNlidar 0.9 mreturn lands behind the wall→ read as the way continuing
CHANNEL CGROUND TRUTH · what is actually there
SECTIONclear distance 3.4 minsulated glass10 + 12 + 10the curtain wall is a solid boundary, indoors beyond it, not a traversable area
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person sees the glare, sees their own reflection, and knows there is a wall in front of them.
Channel B · return · what the sensor receivesMost of the energy bounces away or passes straight through the glass. What little comes back is off the indoor floor behind the wall, so the way ahead reads as still open.
Channel C · ground truth · what is actually thereThe curtain wall is a solid boundary at 3.4 m, and what lies beyond it is indoors, not a traversable area.
Ground truth
Not passable the curtain wall is a solid boundary, indoors beyond it
Perception verdict
Passable no return ahead, or a return landing behind the wall, read as the passage continuing
Mismatch
False-negative obstacle impassable ground called walkable
Consequence
A head-on strike into the curtain wall or a cabinet door. Broken glass, equipment down.
Trigger conditions
High specular reflectance · incidence close to the normal · worst when the glass is clean and unmarked.
WR-220 wheeled inspection severeQR-X4 quadruped severeUAV-M6 UAV severeCL-P2 rail cleaning n/a
SCHEMATIC · SPECULAR PATHhow one beam loses a curtain wallloops
curtain wall sensor specular deflection diffuse off the floor d d where the return lands mirror image of the floor point the plane of the wall: not one return read as passable schematic · proportions enlarged for reading, no reflectance or signal strength computed
  1. emit
  2. deflect
  3. return
  4. land
  5. verdict
This panel is an animation, not a calculation. The path is drawn from spatial position, the timing is slowed down so you can follow it, and not one number on screen has been computed. For the panel that does compute, see DS-04, the cable trench with a cover missing. The mechanism itself is real. Hit a curtain wall at a shallow angle and it behaves close to a perfect mirror: the light does not come back, it folds down to the floor and returns the same way. All the sensor knows is the direction it fired and the total time of flight, so it drops the point along the outgoing direction at whatever distance the round trip measures. That position is exactly the mirror image of the floor point. The wall itself gets no return at all, and so reads as empty space. Only near the normal does enough energy come straight back for the wall to show up.
Red team · how this one is built

Set an approach line in front of the curtain wall and the cabinet doors, close from 6 m at 0.3 m/s, and record where it first brakes. Run it twice, once with the glass clean and once with dust on it.

DS-03 · LOW ALBEDO / DARK ABSORBING SURFACE

A Black Cable Ramp Lying across the Road

A temporary cable crossing ramp, matt black rubber, 120 mm high. The most common temporary obstacle on a site like this.

CHANNEL AAPPEARANCE · what a person sees
ELEVATION · detailblack rubber cable rampmatt asphalt · backlit hours
CHANNEL BRETURN · what the sensor receives
SENSOR RETURN · detaildepth camera 0.9 mdark surface absorbs near-infrared → no return→ taken as ground continuing
CHANNEL CGROUND TRUTH · what is actually there
SECTION · detail120 mmsolid obstacleWR-220 step limit 80 mm
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person sees a black lip across the road, a different colour from the surface, edges plain.
Channel B · return · what the sensor receivesA matt dark surface swallows almost all the near-infrared. Return strength drops below threshold and gets thrown out as noise, so that stretch of road looks like nothing at all.
Channel C · ground truth · what is actually thereA solid obstacle 120 mm high, over the 80 mm step limit of the wheeled platform.
Ground truth
Not passable 120 mm high, above the step limit of the WR-220
Perception verdict
Passable no valid return, read as continuous ground
Mismatch
False-negative obstacle impassable ground called walkable
Consequence
A wheeled platform hits it at full speed. Best case it jams. Worst case it tips, or crushes the cables inside the ramp.
Trigger conditions
Very low surface reflectance · backlit or in shadow · return strength falls with distance.
WR-220 wheeled inspection severeQR-X4 quadruped slightUAV-M6 UAV n/aCL-P2 rail cleaning n/a
Red team · how this one is built

Put black, grey and white ramps of the same spec through the same spot and measure how far out each one is detected. The spread is your margin on this item.

DS-04 · NEGATIVE OBSTACLE / OPEN TRENCH

A Cable Trench with a Cover Missing

One cover left off after maintenance, the trench 0.45 m wide. Negative obstacle is the umbrella term for anything that goes downward, pits, trenches, open shafts, and it is the class inspection misses most often.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONa narrow slot that looks like a shadowcable trench, cover missing
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNlidar 0.9 mone beam lands in the slot → filtered as noise→ read as continuous ground
CHANNEL CGROUND TRUTH · what is actually there
SECTION0.45 m0.80 mQR-X4 crossing limit 0.40 m · WR-220 cannot cross
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesFrom standing height a person sees the trench mouth as an obvious dark slot with clear edges.
Channel B · return · what the sensor receivesThe sensor sits at 0.9 m. Fire at a trench mouth several metres out and angular resolution puts only one or two beams anywhere in the slot, which is easy to filter away as noise.
Channel C · ground truth · what is actually thereThe trench is 0.45 m wide and 0.80 m deep. A quadruped can cross 0.40 m. A wheeled platform cannot cross at all.
Ground truth
Not passable 0.45 m wide, beyond the 0.40 m crossing limit of the QR-X4
Perception verdict
Passable returns inside the slot are too few to form a negative obstacle, read as continuous ground
Mismatch
False-negative obstacle impassable ground called walkable
Consequence
The whole robot drops into the trench. Quite possibly a write-off.
Trigger conditions
Worse the lower the sensor is mounted and the further the target · detection falls sharply below 0.5 m trench width.
WR-220 wheeled inspection severeQR-X4 quadruped severeUAV-M6 UAV n/aCL-P2 rail cleaning n/a
LIVE SECTIONone trench; move a parameter and the verdict flipsdrag to turn
returns landing in the trench
the robot's verdict
This panel stands the DS-04 section upright. The ground is cut along the plane the beams lie in, and section lines, trench outline and ground line are drawn exactly as in the three views above. Every return point on screen is calculated from the three current settings, not drawn in by hand. Beams leave the sensor one at a time at the set angular resolution, and where they land falls out of the 0.45 m width, the 0.80 m depth and the mounting height together. Beam lines are thinned so you can read them; every return point is drawn. Only spatial position is computed here. There is no material reflectance and no signal strength model. The cluster minimum of three is an assumption, and on a real system you would measure the filter threshold on the robot itself. Take angular resolution from 0.8° to 0.4°, or move the trench from 4 m in to 3 m, and the verdict flips back. That is where your margin sits on this one.
Red team · how this one is built

Build the same trench at four widths, 0.25 / 0.35 / 0.45 / 0.60 m, approach each from 3 / 5 / 8 m, and find the cell where detection drops below 90%.

DS-05 · OPEN GRATING / BODY-DEPENDENT SURFACE

The Grating Walkway over a Cable Trench

Nothing makes the relational point better than this one. The same grating panel is fine for a wheeled platform, and a quadruped puts a foot straight through it.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONwalkway grating, looks like groundwalkway over a cable trench
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNdepth camera 0.9 mcontinuous return → read as solid groundthe sensor did not get this one wrong
CHANNEL CGROUND TRUTH · what is actually there
SECTION · detailvoid belowflat bar 30 / gap 40 mmQR-X4 foot diameter 35 mm < 40 mm gap
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person sees an ordinary walkway, steel grating, walkable.
Channel B · return · what the sensor receivesThe return is continuous, so the ground reads as solid and passable. The sensor did not get this one wrong.
Channel C · ground truth · what is actually thereFlat bar 30 mm wide, gap 40 mm. The quadruped's foot is 35 mm across. Smaller than the gap.
Ground truth
Depends on the platform passable for the WR-220; not for the QR-X4, whose foot goes into the gap
Perception verdict
Passable read as solid ground, the same for every platform
Mismatch
Missing criterion nothing was misread; platform dimensions were never in the criterion at all
Consequence
The quadruped drops a foot halfway along the walkway, loses balance, and somebody has to go in and free it.
Trigger conditions
Most dangerous when the gap is close to foot diameter · independent of light and weather, present at all times.
WR-220 wheeled inspection n/aQR-X4 quadruped severeUAV-M6 UAV n/aCL-P2 rail cleaning n/a
Red team · how this one is built

Nothing optical to interfere with here. Build a size series instead: four sections of grating with 25 / 35 / 40 / 55 mm gaps end to end, walk the same quadruped straight through, and record where it first drops a foot.

DS-06 · AIRBORNE PARTICLES / STEAM & DUST

A Plume of Steam at a Storage Container Vent

On a winter morning the vents on storage containers throw off sheets of water vapour. Dust, rain fog and blown grit all do the same thing.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONstorage container venta plume of vapour
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNlidar 0.9 mreturns off particles → read as an airborne obstacle
CHANNEL CGROUND TRUTH · what is actually there
SECTIONclear width 2.6 m, passable throughoutvapour has no substance and bears no load · road intact
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person sees a plume of steam and knows they can walk straight through it.
Channel B · return · what the sensor receivesDroplets and particles scatter the laser straight back, and the suspended point cloud that forms gets clustered into an obstacle.
Channel C · ground truth · what is actually thereThe vapour has no substance and bears no load. The passage is clear the whole way.
Ground truth
Passable no solid obstacle, clear width 2.6 m
Perception verdict
Not passable returns off suspended particles clustered into an airborne obstacle
Mismatch
False-positive obstacle walkable ground called impassable
Consequence
Repeated hard stops, detours, an interrupted task. A UAV is more likely to be pushed into airspace that really is dangerous.
Trigger conditions
Cold air with the container venting · dusty weather · low visibility in rain or fog.
WR-220 wheeled inspection moderateQR-X4 quadruped moderateUAV-M6 UAV severeCL-P2 rail cleaning n/a
Red team · how this one is built

Run a fog machine halfway along the passage at three concentrations and record which one flips the verdict from passable to obstacle.

DS-07 · DYNAMIC RANGE / LIGHT TRANSITION

Entering an Equipment Room Doorway out of Midday Sun

Your eyes take a second or two to adjust on the way in. So does a camera. The difference is the camera keeps moving through those two seconds.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONdark insidemidday sunequipment room doorway · floor level inside and out
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNstereo camera 0.9 mblack after overexposureno texture → no disparity→ marked unknown
CHANNEL CGROUND TRUTH · what is actually there
SECTION2.2 mlevel difference 0 mmfloor level inside and out · doorway 1.8 m clear
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person sees that it is dark inside but knows the floor runs on, and walks in.
Channel B · return · what the sensor receivesAuto-exposure cannot keep up with a difference of several orders of magnitude. The interior is crushed to black. No texture means no disparity, so the whole region gets marked unknown.
Channel C · ground truth · what is actually thereThe floor is level inside and out, 0 mm difference, and the doorway is 1.8 m clear.
Ground truth
Passable level inside and out, 1.8 m clear
Perception verdict
Unknown no texture in the overexposed region, no depth to compute, marked unknown and the robot stops
Mismatch
Verdict withheld neither passable nor impassable was ever stated, and the robot stops in the doorway
Consequence
Repeated stop-start at the door, or the light-dark boundary mistaken for a step and a leg lifted over it.
Trigger conditions
Illuminance differing by orders of magnitude inside and out · midday and around sunrise or sunset · slow auto-exposure convergence.
WR-220 wheeled inspection moderateQR-X4 quadruped moderateUAV-M6 UAV slightCL-P2 rail cleaning n/a
Red team · how this one is built

Put lux meters either side of the doorway, take three illuminance differences at midday, under cloud and at night, drive in at 0.5 m/s each time, and record how far past the threshold valid depth comes back.

DS-08 · SELF-SIMILARITY / LOCALIZATION DRIFT

Twelve Identical Rows of Modules

This one does not fool obstacle detection. It fools localisation. The site repeats so exactly that the robot loses track of which row it is in.

CHANNEL AAPPEARANCE · what a person sees
ELEVATIONevery row looks the samemodule array · rows 5 to 10
CHANNEL BRETURN · what the sensor receives
SENSOR RETURNrow 7row 9point cloud features nearly identical → matched as the same row
CHANNEL CGROUND TRUTH · what is actually there
PLAN5678910actual offset 12.0 mrow pitch 6.0 m, 12 rows in all, layout entirely repeating
the three ground lines now register as one
drag to turn; square on, the three ground lines snap into one
Channel A · appearance · what a person seesA person navigates off the control building in the distance, the road junction and the row plates, and does not lose count.
Channel B · return · what the sensor receivesPoint cloud features on adjacent rows are nearly identical, and loop closure matches row 9 to row 7.
Channel C · ground truth · what is actually thereRow pitch is 6.0 m, so two rows out is a 12 m position error.
Ground truth
Position correct the robot is actually at row 9
Perception verdict
Position wrong matched to row 7, position error 12.0 m
Mismatch
Pose drift the verdict itself is sound; it has just been pinned to the wrong place
Consequence
Inspection data gets filed against the wrong string number and the defect report points at the wrong equipment. Sorting that out costs more than the robot's downtime.
Trigger conditions
More than a few consecutive identical rows · no artificial landmarks · worst after snow, or during construction when ground features are erased.
WR-220 wheeled inspection severeQR-X4 quadruped moderateUAV-M6 UAV severeCL-P2 rail cleaning n/a
Red team · how this one is built

Walk the array end to end twice, once with a different artificial target at rows 3, 7 and 11 and once without, then compare the endpoint drift.

LIMITATIONS

Notes

Before reading
  • This page is a demonstration sample. The site is invented, and every coordinate, dimension and number on it exists to explain the method. None of them belong to a real site.
  • All figures are indicative, not measured. The mechanisms, and which way each misjudgement runs, are qualitative conclusions.
  • DS-04 computes spatial position only, and the cluster minimum of three is an assumption. It is not a pass or fail on any particular perception system.
  • The four platforms match those in the affordance map, so the sheets can be read against one another.
  • For close reading or printing, this page is the original.

We would be glad to talk with you about a specific scene and how it plays out on the ground.