Wrist camera object presence detection without picking

A wrist camera spends its life finding parts for a gripper, but the locate step alone answers a cheaper question just as well: is the part there? Mounted on the flange with no end effector and jogged to one fixed pose, the camera returns a single pass-or-fail boolean your logic can branch on — no motion after the capture, no gripper I/O, no hand-eye calibration. This page walks the build from the bench side: the snapshot pose, teaching the part, setting the score threshold from measured values, the field-of-view math, lighting, and the 20-cycle test that tells you the check is safe to trust.

Can a wrist camera verify presence without moving the robot?

Yes — command one taught snapshot pose, call the camera's locate tool once, and branch on its pass-or-fail result: a presence check needs only that boolean, so the camera can hang on the flange with no gripper attached and the robot never moves after the capture.

Teach-pendant vision works this way by design: the camera-locate step runs its child branch only when it finds the taught part, so a one-line assignment stores the outcome in a boolean your program reads afterwards. A smart camera does the same inside its inspection job — the pattern-match tool returns a score and a position, and for presence you consume the score alone.

Because you never use coordinates, hand-eye calibration drops out of the project. Calibration exists to convert pixel positions into robot-base coordinates; a presence flag does not touch them. Add a solve only when the same snapshot must also feed a pick position — the smart camera hand-eye calibration guide covers the case where no calibration wizard exists.

What do you need before you teach the presence check?

You need a repeatable snapshot pose, a taught reference image of the part under fixed light, and a locate tool with an adjustable score threshold — plan half a day at the bench, where a camera stand is easier to work on than the robot itself.

• A camera with a locate tool. One still frame per check is all the matcher needs, so a 1.6 MP class sensor is plenty for presence duty — the worked example below shows why. The MV-SCB007EC-12S-WBN is that class: a 1.6 MP color vision sensor with a 12 mm manual-focus M12 lens on Fast Ethernet.

• A fixed snapshot pose. The bracket holds the sensor square to the checkpoint; you record joint positions once.
• Controlled light. Integrated LEDs at a locked level, or an added diffuse light for glossy parts.
• A repeatable fixture. The taught image has to cover the part's placement tolerance; the tighter the fixture, the wider the score margin.
• A bench to build on. Prove the recipe before it touches the robot: the LS-MNT-0040 stand holds a fixed 700 mm working height over a 400 × 450 mm base, enough room to set distance and lighting without tying up the cell.

How do you build the check, step by step?

The build is four moves — pose, light, teach, branch — and the whole cycle fits inside a second on a smart camera, because nothing moves except the one jog to the snapshot pose.

1. Command the snapshot pose. Jog the flange to roughly 250 mm above the checkpoint, sensor axis square to the fixture, and record the joint positions. With no gripper there is no tool frame to define; allow about 200 ms of settle time after motion stop before any capture.
2. Lock the light. Set the sensor's LEDs to a fixed level and mask any daylight reaching the checkpoint.
3. Teach the part. Capture at the pose and draw the taught box on the most distinctive stable feature, keeping at least 30 pixels across it. Note the teach score; a clean teach lands at 0.95 or above.
4. Measure the empty case. Capture five times with the checkpoint empty and record the worst false score.
5. Measure the loaded case. Capture five times with the part placed at the worst allowed offset and rotation, and record the lowest true score.
6. Set the threshold midway between those two numbers, inside the 0.5 to 0.8 band — the next section works a full example.
7. Program the branch. In the locate step, success runs the child branch: set part_present — a Boolean, initialized false — to true there, to false in the failure branch, and log the score every cycle. This is the same flag pattern the vision locate with a part detected flag guide uses to stop an empty belt.
8. Verify with the 20-and-20 test in the verification section before trusting the check on the line.

What score threshold decides a part is present?

Set the accept threshold between 0.5 and 0.8 on the 0-to-1 scale and choose the exact value as the midpoint of your two measured extremes: a typical bench run pairs a 0.98 teach score with a 0.52 worst false match, and 0.65 splits them.

Both example scores are the kind of spread a bench run produces; measure your own rather than copying them. Pattern-match scores are normalized correlations: a perfect match to the taught image scores near 1.0 and everything else falls below it. The 0.5 to 0.8 accept band is the working range most locate tools need — below roughly 0.5 a match survives heavy degradation, above roughly 0.8 in-tolerance parts begin to miss.

Threshold = (0.78 + 0.52) / 2 = 0.65. That leaves a 0.13 margin above the worst false match and 0.13 below the worst true score. Keep at least 0.05 of margin on each side; if the two distributions overlap, no threshold separates them — fix the contrast or teach a second feature instead.

Bias toward the empty side when the margins are unequal: a false reject stops the line, a false accept lets the error travel downstream. The pattern match score threshold guide explains how the coarse search gates which candidate gets scored — worth reading before you chase a low score with threshold edits.

How much field of view does a presence checkpoint need?

Size the field of view to the checkpoint plus a 20-25% margin, then hold at least 5 pixels per millimeter so the taught feature keeps 30 or more pixels across it — enough headroom for the matcher without wasting sensor area.

Example: the checkpoint is an 80 mm pocket in a tray, so with a 25% margin the field must be 100 mm wide. Assume a 1/2.9-inch-format 4:3 sensor, about 5 mm across, at a 250 mm working distance: f = 5 × 250 / 100 = 12.5 mm, so a 12 mm lens gives FOV = 5 × 250 / 12 ≈ 104 mm. A 1.6 MP 4:3 sensor resolves about 1460 × 1095 pixels (√(1,600,000 × 4/3) ≈ 1460), so px/mm = 1460 / 104 ≈ 14, and an 8 mm taught feature spans 8 × 14 ≈ 112 pixels — nearly four times the 30-pixel minimum. The same checkpoint on a 0.3 MP VGA sensor gives 640 / 104 ≈ 6 px/mm, so the feature spans about 49 pixels: presence still passes; precise position measurement would not.

Presence checks can run on far less sensor than that first case: the MV-SCA002M-03S-WBN-SR is a 0.3 MP mono vision sensor with a 3 mm lens focused at 240 mm, made to sit at one fixed distance over a checkpoint.

Run your own checkpoint width, sensor and distance through the focal-length calculator below.

What lighting keeps a presence check stable?

Keep the light at judgment time identical to teach time: lock the integrated LEDs to a fixed level, mask daylight, and add a diffuse ring or a polarizer for glossy parts, because lighting drift between teach and run is the top cause of false results.

Ambient daylight is the quiet killer: a dock door opening between teach time and run time shifts every score in the job. Shield the checkpoint, and re-measure the score spread after any change to surroundings — the ambient light variation guide lays out the control strategy for pick-and-place cells.

Keep the matcher on mono unless color is the only difference between present and absent. Color checks drift when the illumination spectrum changes; if you must detect by color, plan on periodic re-teaching — this guide to fixing color-based detection on wrist cameras covers the usual repairs.

Two sensors in the rail attack lighting from opposite ends. The MV-SC6050C-08M-XBN carries an integrated multi-color light source around its 5 MP color sensor, so the illumination travels with the recipe. The mono MV-SC5060XM-08M-WBN on GigE keeps color out of the score entirely, and its rolling shutter costs nothing while the robot holds still.

When is a fixed overhead camera the better answer?

Go fixed-mount when the robot's time is worth more than the camera's: an overhead area scan camera checks the station on every frame with no robot motion at all, while the wrist camera wins when one sensor must serve several poses by jogging between them.

A fixed camera also removes the settle-delay problem outright: nothing moves during its capture. For that role, a camera such as the MV-CU020-90GC — a 2.5 MP global-shutter GigE model on a C-mount — watches the station permanently and leaves resolution to spare for position checks later.

How do you verify the check before you trust it?

Run 20 cycles with a part present and 20 with the checkpoint empty, log every score, and ship the recipe only with zero false accepts and at least 19 of 20 true accepts — a score within 0.05 of the threshold means the margin is too thin.

Log the score, not just the flag: score history shows a margin shrinking before a miss happens. Add two negative tests. The nearest look-alike part from another product must fail. A part missing its decisive feature — no hole, no print — must fail if that is what the check exists to catch.

Re-run the 20-and-20 after any change to lighting, lens, bracket or fixture, and after the line has run hot and cold for a full day. Decide the failure action before go-live: one automatic re-capture clears transient glare; a second miss stops the cell and raises the alert.

When does a wrist-camera presence check fail?

Expect failure in three cases: the part sits in an orientation you never taught, glare or transparency hides the taught pattern, or the capture happens before the robot finishes settling — each has a bench test and a fix cheaper than reworking the line.

• Untaught orientation. A part rotated 90° from the taught view reads as no part at all. Teach each orientation as its own model, or constrain the delivery fixture so orientation cannot wander.
• Glare and transparency. Tilt the sensor 15-30° off the specular direction and add a polarizer; for transparent parts, light the checkpoint from behind instead of the front.
• Motion and focus. Capture only after the settle delay, and remember a manual-focus lens drifts as the cell heats and cools — the camera focus drift guide shows daytime blur arriving first as score decay.
• Small features. Parts under about 5 mm strain teach routines; this small-part teach failure guide walks the standard fixes.
• Not enough viewpoints. When one pose cannot see the whole checkpoint, add snapshot poses rather than widening the lens — the multiple snapshot positions guide covers serving a large table from one camera.

And the honest boundary: if presence is the only question, the checkpoint is single, and there is no wrong-part risk, a photoelectric sensor is the cheaper instrument. Choose vision when the check must tell the right part from the wrong one — the score does that; a light barrier cannot.

Shortlist sensors in the smart cameras category or the camera finder, then size distance and lens in the camera configurator. For a quote on a multi-camera line, contact us.