Ambient light variation in pick and place vision: control it

A vision cell that picks cleanly in the assembly plant and misses after the move to a brighter hall does not have a robot problem; it has an optics problem. Room light reaches the part, adds to your own illumination, and every grey level the vision models were taught on moves with it. Below are the numbers behind that drift: how much a stop of extra ambient light shifts the image, which vision tools fail first, and the three fixes that hold — shielding the station, strobing your light, or moving off-band with IR illumination and a pass filter. It closes with the acceptance test that proves the cell survives its next room, whichever fix you choose.

Does a lighting change really break a pick and place vision system?

Yes — room light that reaches the part adds to your image exactly as your own light does, and a cell tuned at 300 lx can meet 1,200 lx or more under exhibition or skylight lighting, which moves every taught grey level.

The signature is always the same. Golden parts pass at the factory; at the new site you see centroid offsets, blobs that drop in and out, or match scores that fall just under the accept threshold. The robot still repeats to microns, so the geometry has not moved — the image has.

Rank the causes by how often each turns out to be the guilty one:

1. Ambient illuminance on the part changed — relocation, extra hall lamps, seasonal daylight through roof lights. Even a factor of four rewrites the image.
2. An auto mode is chasing the scene — auto exposure or auto gain re-balances every frame and drags your taught grey values, which is why every auto mode goes off.
3. Mains-powered hall lamps flicker at 100 or 120 Hz, and exposures not locked to the mains give frame-to-frame brightness scatter.
4. Daylight patches sweep across the field as sun and clouds move — non-uniform, and the hardest case to teach around.

How much does ambient light move the grey levels in the image?

Grey levels from the room scale with the illumination on the part: double the ambient and the room's contribution to every pixel doubles, so a 40 DN offset at 300 lx becomes 80 DN at 600 lx — one stop of light, a factor of two in DN.

Sensor signal is proportional to the light on the part, times the exposure, divided by the f-number squared. Your light and the room's light add at the scene and then share the same exposure and iris, so only their ratio at the part decides how far thresholds can drift.

Worked example — the exhibition move. In the workshop, 300 lx reaches the station, the camera runs a fixed 5 ms exposure at f/4, and your flood light drives the dark part to 30 DN and the conveyor background to 170 DN. With your light off, the room alone reads 40 DN. Totals: part 70 DN, background 210 DN, threshold 120 DN — 50 DN of margin each way. The fair's lamps lift the station to 1,200 lx. The room's share becomes 40 × (1,200 / 300) = 160 DN, so the part now reads 30 + 160 = 190 DN, above the 120 DN threshold: the dark blob is gone, and the background clips at 255. Nothing mechanical changed; only the room's term did.

Stopping down to f/8 or cutting exposure to 1.25 ms divides your light, the background and the room's light by the same factor of four, so the ratio — and the failure — survives untouched. That is why every fix below changes the lights or the optical path, not the camera settings.

Set the budget before changing anything: keep the room's share under 10% of your working signal, about 20 DN on an image that swings 200 DN. That caps the worst single move at a fifth of a stop and leaves your thresholds their full margin.

Which vision tools fail first when the light changes?

Blob and fixed-threshold tools fail first, grey-level correlation templates next; shape- and edge-based matching models hold out longest because they score on geometry, not absolute grey.

In order of how soon each gives trouble:

1. Fixed-threshold blob tools — thresholds are absolute DN values, so a shift like the 120 DN above is fatal. They fail first, every time.
2. Grey-level correlation and normalized templates — contrast normalization absorbs a stop or two of uniform change; saturation removes the texture they match.
3. Contour- and shape-based matching models — they score on edge geometry and normalized contrast, and hold until the edges themselves wash out in clipping.

Run a five-minute test before touching hardware: replay your saved golden images with pixel values scaled up by one and two stops, and record each model's score. The tool whose score collapses first is the one the ambient will take down first — replace it, or make its light independent of the room.

Should the camera's auto exposure absorb the change?

No — auto exposure re-aims every frame at mid-grey by trading exposure time against gain, so the grey scale your thresholds were taught on moves by design; fix the exposure and gain and control brightness at the light.

On a strobed cell auto modes are worse than useless: the metering sees a mostly dark frame between pulses and drives exposure up until the pulse saturates the sensor. Set exposure, gain and light level as fixed values instead. If one auto mode must survive for setup, cap its maximum exposure time so it can never blur a moving part. Contour position jitter under automatic exposure is the usual symptom that one was left on — the edge wanders tenths of a pixel while the exposure hunts.

How much does a shield or enclosure cut the ambient?

A fitted shroud around camera and part removes most of it: with matte interior walls and a part window facing away from the lamps, expect the room's contribution to fall by 10-30×, and measure what is left with your own light switched off.

The standard answer in guided-pick cells is a light hub: a box around the camera and the part that carries its own diffuse light. With a white matte interior and one or two diffuse sources inside, the part sees even illumination and the room sees only the box; a matte-black shroud without internal light is the lighter version. The entrance window is the only leak — face it away from the lamps and toward the floor.

The rejection is geometry, not a coating: the camera and part now see enclosed surfaces instead of the room, and the range above assumes a window that stays small against the box. Measure the actual number. Lock exposure and gain at the working values, capture with your light off, and read the mean grey — that is the room's share in DN. Switch the light on and confirm the working signal puts it under your 20 DN budget. Power the inside light through a light extension cable routed through a grommeted hole.

The cost is reach and access: the gripper must enter the box, and the box must open for maintenance. For small parts under an overhead camera it is still the fastest fix.

Does strobing the light really improve the light-to-ambient ratio?

Yes, by a factor you can compute: delivering the same part photons in a 0.5 ms pulse at 10× overdrive instead of a 5 ms flood gives the room 10× less time to expose the sensor — the rejection factor equals the overdrive factor.

Photons from your light scale with intensity times on-time, so a pulse at ten times the current for a tenth of the time delivers the same part signal. The room's light only integrates while the sensor is exposed, so it falls with the pulse time:

Applied to the exhibition example: replace the 5 ms flood with a 0.5 ms pulse at 10× overdrive. The room's share becomes 160 × (0.5 / 5) = 16 DN, so the part reads 30 + 16 = 46 DN and the background 186 DN; the 120 DN threshold now separates them with 74 and 66 DN of margin, in either room. Motion blur falls with the pulse too: a belt at 300 mm/s smears 300 × 0.005 = 1.5 mm under the flood, but 0.15 mm under the pulse.

Two integration details decide success. The exposure window must overlap the pulse — a strobe firing outside the exposure gives a black image, and the fix is trigger timing. The pulse current must also stay inside the light's rated pulse current and duty cycle; check that against the light's data before choosing the controller. How to measure the before-and-after rejection properly is its own procedure: see measuring strobed lighting ambient rejection.

Any digital controller with a strobe feature can run this scheme:

The LS-CTL-0070 is a four-channel digital control box rated 200 W, with RS-232 and a strobe feature. The LS-CTL-0264 is a high-power digital controller, two channels at 400 W, with a network option. The LS-CTL-0180 is a constant-current controller, four channels at 120 W, for 24 V constant-current lights. Match channel count and output type to the light, then browse the light controllers or the controller finder filtered to strobe-capable models.

When is IR light with an IR pass filter the better fix?

When the station cannot be boxed in: an 850 nm LED plus a matching band-pass filter on a mono camera passes your own light at around 90% at the LED wavelength while blocking the visible room light by three to four orders of magnitude (OD 3-4).

The camera sees your 850 nm LEDs through the filter; LED and fluorescent hall lamps emit almost nothing beyond 700 nm, so a visible-lit hall collapses to a few DN of leakage. The method loses its power wherever the room itself is rich in infrared — halogen and incandescent lamps and daylight all carry strong IR — so under tungsten light or outdoors, shield or strobe instead.

Two requirements come with it. The camera must be mono: a color sensor's Bayer mosaic and IR-cut filter pass only a few percent of 850 nm, throwing away most of the margin. And focus shifts with wavelength, so focus with the IR light running, or stop down until the shift sits inside the depth of field.

The HK-TD032-12-WBN is a 1.6 MP mono vision sensor with a global shutter and an M12 lens mount, a compact fit over a pick station.

Where you would rather screw the band-pass filter into the lens itself, use a C-mount camera such as the LS-SMC-0017, an x86-platform smart camera with an 8.8 MP sensor, global shutter and C-mount.

Filter the smart cameras by mono, global shutter and mount, or open the camera finder.

What acceptance test proves the cell survives the next room?

Reproduce the extremes before shipping: run the cell with your light off, hall lights full on, hall lights off, and a bench lamp aimed at the station, and require the taught position to move less than a fifth of your pick tolerance.

Run the same five steps at the factory and again on site:

1. Locked settings. Exposure, gain and light level fixed; every auto mode off.
2. Ambient-only capture. Own light off at working exposure; the mean grey must sit under your 20 DN budget — 10% of a 200 DN swing.
3. Four-room sweep. Hall lights on, hall lights off, bench lamp into the station, daylight if a window is near; 30 pick cycles per condition.
4. Pass rule. Position repeatability within a fifth of the pick tolerance. At an example scale of 0.03 mm per pixel (a 60 mm field across a 2,000-pixel sensor), a 0.05 mm tolerance demands better than 0.01 mm — about a third of a pixel, ordinary centroid repeatability.
5. Record it. Log model score and position per condition, and re-run the whole sweep after any hall lighting change.

Put the numbers into the acceptance specification so the next relocation is a test, not a surprise; the ambient-light rejection specification article shows how to word it, and the lighting stability test covers the pre-deployment checks.

Shield when you can, strobe when you cannot enclose, and go IR when the hall is LED-lit and the station must stay open. To size the controller and light for the overdrive you settle on, start at the lighting configurator or send part photos through the contact page.