Reduce wrist camera snapshot time: fix focus and exposure

A wrist camera locate step that costs 7-8 seconds per pick is usually not short on computing power. The camera is repeating its own setup on every shot: autofocus sweeps the lens motor, auto-exposure shoots test frames, auto white balance re-balances a color sensor, and the light driver re-syncs with the trigger. Each of those costs whole seconds before the frame you actually use is taken. The fix is mechanical. Time the step, fix focus, exposure, white balance and light control at taught values, teach in CAD-guided mode where the software offers it, and confirm the locked focus still holds across the arm's travel — the worked budget below takes a 7.5 s locate down to about 2 s.

Read the telltale first: the same seconds everywhere

Start with the stopwatch, not the algorithm. If the locate takes about the same time at every position — 7.5 s at the far snapshot, 7.5 s at the close one — the cost is fixed overhead, not a hard scene. A scene-driven cost moves with the part: a cluttered bin takes longer to search than an empty one. A setup-driven cost is flat, because the camera runs the same test-and-adjust sequence no matter what sits in front of the lens.

Four automatic functions produce that flat overhead:

• Autofocus steps the focus motor, scores image contrast after each step, and repeats until contrast peaks. Each step costs at least one frame plus motor settle. At a 15 fps live stream a frame costs 1/15 ≈ 0.067 s, so a 20-step sweep with 0.03 s settle per step runs about 20 × (0.067 + 0.03) ≈ 1.9 s before the first usable frame. Your own stream rate and the camera's timing output turn that estimate into a measurement.
• Auto-exposure shoots a test frame, reads the brightness and applies a new exposure time — at least one extra frame and settle per shot.
• Auto white balance re-measures a color sensor's gray balance on a similar test cycle. Mono sensors never pay it.
• Automatic light control — when the camera commands the light per shot, the command round-trip adds its own fixed slice to every cycle.

If you run an arm-mounted camera kit, its wrist camera scan-time troubleshooting notes point at the same switches.

What you need

• A stopwatch or the cell cycle timer, and ten consecutive locate cycles to time.
• Access to the camera settings inside the snapshot step of your vision software: focus mode, exposure mode, white balance mode, light mode.
• The real part in its real fixture, and the CAD model if your teach supports it.
• A tape measure or shim set, to measure how far the working distance wanders between picks.
• The camera's live stream frame rate, so autofocus steps convert to seconds.
• The depth-of-field calculator, for step 3 below.

Cut the snapshot time, step by step

1. Time the baseline. Run ten locate cycles and divide the total by ten. Record the far and close positions separately. Two nearly equal numbers mean setup overhead, and the rest of these steps recover most of it.

2. Switch focus to manual first. It is usually the largest single cost and the safest to change. Focus once at the real working distance with the part in place, and write the value down. Every cell can flip this switch; very few cells truly cannot.

3. Check the depth of field before you trust the lock. Manual focus holds only while the part stays within the depth of field around the taught distance:

Here f is focal length, WD the taught working distance, N the f-number and c the circle-of-confusion criterion — take c as 0.02 mm, a few pixels on a typical small sensor. Example: a 12 mm lens at f/2.8 focused at 300 mm gives m = 12/(300 − 12) = 0.0417, so DOF ≈ 2 × 2.8 × 0.02 × 1.0417 / 0.0417² ≈ 67 mm, i.e. ±33 mm. If your worst-case distance spread sits inside that window, the lock is safe. Stopping the lens down one or two f-stops widens the window roughly in proportion. This is an estimate, not a datasheet value — bracket it by picking from the highest and lowest part positions and comparing lens-to-part distances with a shim.

4. Fix exposure: lock exposure time and gain together. The arm is stationary during a snapshot, so motion blur is rarely the limit. If you shoot on the move, blur = speed × exposure time: at 100 mm/s with a 5 ms exposure that is 100 × 0.005 = 0.5 mm, which must stay under about a third of the smallest feature you resolve. Inside the lock, add exposure time before gain — gain lifts noise with the signal, and if the locked image comes out darker than the auto one, that is usually the reason: why auto exposure beats a gain-heavy manual setup.

5. Lock white balance on color cameras. Teach the balance once under the production lights and re-lock it after any lighting change. On a mono sensor this step does not exist, which quietly removes another slice from every shot.

6. Fix the light. Run the light constant, or strobe it from the same trigger that fires the camera, with the duration set once in the job. A vision controller does this in hardware: the MV-VC2140-128G40-1T carries four built-in light channels, so each channel's mode and duration belong to the job rather than to a per-shot negotiation.

7. Teach with CAD guidance where the software offers it. Automatic teach scans the physical part to build its locate model; CAD-guided teach starts from the model you already have, so teaching is quicker, and re-teaching after a design change is a model update instead of a re-scan. If you configure locate inside an arm-mounted camera plugin, the walkthrough on configuring camera locate snapshot positions sets up the same snapshot positions step by step. Keep those positions honest with stream versus snapshot mode — that distinction is where locates pass at teach and fail in production.

8. Re-time after each change, one switch at a time. Lock focus, run ten cycles, note the number. Then exposure, then white balance, then light. The stopwatch tells you which switch paid, and the total should land near the budget in the next section. If the settings vanish after a power cycle, make the job load itself at start — see auto-loading the vision job on startup.

The time budget: where 7.5 s goes

Your stopwatch gives the total; a per-function split turns it into a plan. The 7.5 s total is the measured figure; the split below is a typical planning value — refine it with the camera's timing output as you switch each function off.

The arithmetic: 7.5 − (2.5 + 1.5 + 1.0 + 0.5) = 7.5 − 5.5 = 2.0 s per locate. At 1,000 picks per day that is 5.5 × 1,000 = 5,500 s, about 92 minutes of robot time returned. A cell that locates twice per cycle — far approach, then close confirm — saves 2 × 5.5 = 11 s per cycle: a 30 s cycle drops to 19 s, lifting throughput from 3,600/30 = 120 to 3,600/19 ≈ 190 cycles per hour.

When the camera is still the bottleneck

Once the four switches are off, the floor is image size, transfer and the locate model itself. Cutting resolution, region of interest or model complexity pays from here: Speed up vision acquisition time by cutting processing load works that side of the problem, and where to set the ROI decides how much image the algorithm sees at all.

If the hardware is up for renewal anyway, choose sensors whose focus cannot hunt. The MV-SCB007EC-12S-WBN pairs a 1.6 MP color global-shutter sensor with a 12 mm manual-focus M12 lens: focus is set by hand and stays set.

The MV-SCA002M-03S-WBN-SR removes the mechanism entirely — a 0.3 MP mono sensor behind a 3 mm lens fixed at a 240 mm focus distance — so there is no sweep to switch off at all.

When one snapshot must carry more parts or a heavier model, step up in resolution rather than in setup automation. The MV-SC3050XC-16M-ATN-H reads 5 MP in color with a global shutter over Fast Ethernet; the MV-SC5060XM-08M-WBN steps up to 6 MP mono on GigE.

Checks before you lock the values for good

• Depth of field against distance spread. Re-measure the spread after any mechanical change — new gripper, new fixture, new part revision. The near and far limits from step 3 must bracket it.
• Blur at the fastest wrist speed. Recompute blur = speed × exposure time whenever you shoot on the move or lengthen exposure for a dimmer cell.
• Ambient stability. A locked exposure assumes the scene brightness stays put; sunlit hours and overhead fixtures move it. The ambient light control fix for pick-and-place vision addresses that side, and an image that slowly brightens or darkens between shifts usually means an auto mode survived somewhere in the job — hunt it down with the same one-switch-at-a-time timing you used in step 8.
• Settings survive power-up. Re-run the timing after a cold start, not only after a warm edit; a device that reloads defaults quietly puts the seconds back.

Common mistakes

• Tuning the algorithm first. A flat 7.5 s is not compute; the four switches come before any ROI or model surgery.
• Focusing on a tilted part or at the wrong stack height, then trusting the lock everywhere.
• Locking exposure under office lighting and running the cell on a dim shop floor.
• Changing two settings between timings and losing track of which one paid.
• Specifying color for a geometry-only locate, then paying the white-balance step on every shot.

When keeping autofocus is the right call

Do not fix focus when the distance spread is wider than the depth of field and the mechanics cannot tighten it — a deep bin, a long reach, mixed part heights. Keep autofocus and recover the rest: exposure, white balance and light control were worth 1.5 + 1.0 + 0.5 = 3.0 s of the 5.5 s in the budget, so the locate lands near 7.5 − 3.0 = 4.5 s instead of 7.5 s. Trim the sweep too: fewer focus steps, or a focus check against a fixed distance reference at shift start, cuts the sweep without removing it.

Set it up and re-time

Run the numbers on your own cell first; the stopwatch is the whole business case. When you spec the hardware, the smart cameras and vision sensors category filters by mount, resolution and shutter, the camera finder narrows by interface, and the vision configurator starts at the camera step and walks you through a matched light and controller. More cycle-time fixes sit on the systems topic page, and if a timing log resists the budget above, bring it to the contact page.