Pick rate for vision guided part presentation on a feeder

A quoted pick rate is where feeder cells are won or lost, and it is decided before the robot moves: by how many separated parts the camera sees in one image. This guide is for anyone adding vision guided picking to bulk parts on a vibratory or flexible feeder. You will get the cycle formula, a worked 30 parts-per-minute example, the pixels-per-millimeter calculation behind the camera choice, and the backlight and trigger decisions that hold the rate up in production.

What sets the pick rate of a vision guided feeder cell?

The pick rate is set by how many pick candidates the camera can see at once in its ROI and by the robot's time per pick, so five candidates per image at about 1.5 s per pick lands near 30 parts per minute. Three numbers decide it: N, the count of candidates one image shows; t_pick, the robot's time per pick; and t_vibrate, the feeder run that brings the next batch into view. Exposure and locate time sit inside t_vision and are usually the smallest term, a few tenths of a second on a vision sensor. Robot motion dominates: about 1.5 s per pick is realistic for a cobot moving 300-500 mm between plate and outlet, while fast arm-class robots reach 0.6-0.7 s. The presentation hardware exists to raise N, and N multiplies through the cycle: five candidates per image means the feeder vibrates once per five picks instead of once per pick.

How do you calculate parts per minute from candidates per image?

Use rate = 60 × N ÷ (t_vision + N × t_pick + t_vibrate): with five candidates, a 0.5 s image, a 1.5 s pick and a 2.0 s vibrate, the cell runs 60 × 5 ÷ 10.0 = 30 parts per minute. The robot works through the candidate list from one image, re-images without vibrating, and pulses the feeder only when the count reaches zero.

In the table, ppm is parts per minute.

Sustained rates near 70 ppm need about ten candidates, a 0.65 s pick and a snappier 1.5 s vibrate: 0.5 + 10 × 0.65 + 1.5 = 8.5 s, so 60 × 10 / 8.5 ≈ 71 ppm. That is why candidate count is the first variable to engineer, not robot speed alone.

Why does a backlit feeder beat a bin for candidate count?

A feeder spreads parts into a single layer over a backlight, so the locator sees several separated silhouettes per image, while a bin shows overlapping parts that count as one candidate and lock the cell near 15 parts per minute. A pile merges touching parts into one blob, so N collapses to one no matter what the robot could do. A feeder plate spreads parts into a monolayer, and a backlight under the plate turns every part into a high-contrast silhouette with clean edges the tool cannot miss. Two refinements raise N further. Plates with molded pockets or channels park each part in a known pose, so the check becomes presence or absence at fixed positions and the robot already knows the grip. Counting parts as separated rather than merged takes a few pixels of clear background between silhouettes — the same clearance problem the gripper has, covered in checking minimum clearance between parts before a robot pick. For the cell layout around a feeder, start with vision-guided flexible feeding for random picks; detecting many parts in one snapshot shows the multi-candidate locate step. The same presentation tuning is documented in robot-side guides such as configuring part presentation for robot scanning.

What field of view and pixels per millimeter does the plate need?

Match the field of view to the plate and hold about 7 pixels per millimeter: a 1.6 MP sensor over a 150 mm plate gives 1080 ÷ 150 = 7.2 px/mm, enough for a 20 mm part that then spans 144 pixels.

Take a 1.6 MP sensor in 4:3 format as about 1440 × 1080 pixels over a 150 mm plate: 1080 / 150 = 7.2 px/mm. A 20 mm part spans 20 × 7.2 = 144 pixels, and a 2 mm feature the locator can key on spans 2 × 7.2 ≈ 14 pixels, both comfortable. The 3-pixel rule sets the floor: 3 / 7.2 ≈ 0.4 mm is the smallest feature worth asking the tool to find. The same arithmetic sets plate size for a bigger sensor. A 5 MP sensor in 4:3 has about 2592 × 1944 pixels, so at the same 7.2 px/mm the short axis covers 1944 / 7.2 = 270 mm, and over a 300 mm plate it still delivers 1944 / 300 ≈ 6.5 px/mm, resolving down to 3 / 6.5 ≈ 0.46 mm. Candidates per image is therefore mostly a field-of-view decision. Run your own plate through the resolution calculator below, and choosing camera resolution for dimensional inspection works the same numbers when the check measures rather than locates.

Which backlight suits a feeder plate?

A red flat backlight sized to the plate, mounted directly under it, gives the strongest silhouette contrast per watt; white suits a color camera that also reads top-surface features, and 54 W over 300 x 200 mm is a typical continuous rating. Size the panel to the plate or slightly larger, and choose wavelength by camera and cycle: red suits mono cameras and efficient LEDs, white suits color cameras, and 850 nm infrared suppresses ambient light and visible glare. Keep the panel dark during vibration so the exposure sees settled parts. Constant-on is fine when the robot images a still plate; strobing freezes motion, lets the drive push higher peak current and cuts heat. A center hole in the panel passes a fixture post through the lit area.

The LS-BKL-1468 red panel at 150 x 150 mm suits small plates, the LS-BKL-1397 covers 300 x 200 mm in white with a center hole for a fixture, and the LS-BKL-0865 is the 400 x 300 mm high power panel rated 92.2 W for strobed cycles.

Where the cell has no light controller, a driverless constant-on panel is the simple path: the LS-BKL-0416 runs 850 nm infrared, constant on only, over 136.4 x 136.4 mm. If the task is measuring edges rather than finding parts, collimated panels buy straighter edges — collimated or diffuse backlight for edge gauging compares the two. The backlights category holds the full range, and the lights finder filters it by size, color and wavelength.

How does the vibration trigger loop work with the robot?

Wire a robot digital output to the feeder's vibration input and pulse it only when the candidate count is zero: one 2 s pulse per batch, not between every pick, is what turns a 15 parts-per-minute cell into a 30 parts-per-minute cell. Run the loop like this:

1. The robot requests an image with the vibration off and the plate settled.
2. The camera returns a candidate list, or an empty result with count zero.
3. Count above zero: the robot picks from the list, then re-images without vibrating, until the count reaches zero.
4. Count at zero: pulse the feeder for about 2 s, wait roughly 0.5 s for parts to settle, and re-image.
5. Three empty pulses in a row: raise a refill request and stop rather than cycling on an empty plate.

Imaging a still plate is what keeps silhouettes sharp, and a global shutter sensor tolerates the residual motion of parts just settling. When parts flow past a fixed point instead of settling, a photoelectric sensor can pre-trigger the exposure at the right instant; photoeye triggering for vision pick rate works through that variant. If results ever lag the plate state, triggering a fresh measurement per pick is the fix for stale-image failures.

Smart camera or vision controller for a feeder cell?

A vision sensor or smart camera handles locate-and-count on the device and returns the pick list to the robot over Ethernet, so you add a vision controller only for strobing the backlight from built-in light channels, extra cameras or central recipe and result storage. On the device path, the camera does the locating and the robot holds the logic, which is the simplest wiring and the fastest to commission. A controller earns its slot when the backlight should be strobed from a light channel timed to the trigger, when a second camera joins the cell, or when recipes and results should live in one place. The MV-VC3301H-128G66, for instance, carries 4 built-in light channels and 5+ LAN ports on an Intel Core i3 with 8 GB of memory, with its VM6200Pro license included — see the vision controllers category.

On the camera side the choice is pixel count and shutter:

All 2 share: Sensor type CMOS, global shutter; Pixel size 3.45 μm × 3.45 μm; Data interface Fast Ethernet (100 Mbit/s); Humidity 20% RH to 95% RH (no condensation); Client software SCMVS; Certification CE, KC. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The 1.6 MP mono sensor covers a 150 mm plate at 7.2 px/mm by the arithmetic above and is enough for presence, count and locate. Step up to 5 MP when the plate grows toward 270-300 mm or the tool must key on finer features. Both sensors here are global shutter on Fast Ethernet, so image timing is deterministic against the trigger. Browse the smart cameras range or filter by interface and resolution in the camera finder.

When is a feeder with 2D vision the wrong answer?

Feed-and-backlight 2D is the wrong answer when parts tangle or nest so vibration cannot singulate them, when the decision needs the part's top surface rather than its silhouette, or when you need more than x, y and rotation to pick. Tangled parts — springs, stamped leads, rubber seals — never separate, so N stays near one and no lighting fixes it; a bowl or magazine with orientation features does that job. If accept or reject lives on the part's top face, silhouette contrast is the wrong tool and the part needs directed top lighting. If the grip needs height, tilt or a 6-axis pose, a 2D plate cannot supply it — when to choose a 3D machine vision system sets out that boundary, and 3D bin picking for flat parts is the closer fit when parts must come straight out of a bin. On a plate, the locator still has to survive part rotation: setting up vision locate to survive rotation covers the robustness checks.

What do you check before ordering the parts?

Before you order, confirm four numbers: field of view covers the plate with margin, the locator resolves the smallest feature at 3 pixels or more, the candidate count in a test image is five or more, and the cycle budget closes at your target rate. Work down this list:

• Field of view: the plate fits the sensor's field with 5-10% margin at the real working height.
• Pixels: px/mm from the formula above; the smallest feature the locator must key on spans 3 pixels or more.
• Candidates: a test image of the loaded plate holds five or more separated parts, or the rate target needs replanning.
• Lighting: panel size covers the plate, wavelength matches the camera, and strobe wiring exists if parts ever move during exposure.
• Trigger wiring: one robot output to the feeder, one camera result to the robot, and a settle time measured, not guessed.
• Cycle budget: the formula closes at your target ppm with t_pick measured on the actual robot path.

Send the plate size, the part and the rate target through the lighting configurator, or ask the team through the contact page — the camera, backlight and controller choices above then come back as a checked parts list. The applications topic page collects the rest of the picking guides on this site.