Fixed or robot-guided drill bit picking: what the cost jump buys

You already have the 6-axis arm and a vision bench, and the job is metal drill bits: 0.5 mm to 13 mm metric plus the inch series from 1/16 in to 9/64 in (1.6 to 3.6 mm), nothing over 1 kg, no sorting required — but every tip must show its 118° point when picked and again when placed. The claim you heard is correct: bits standing end-up in a known nest keep this a one-camera job, while letting the arm find loose bits turns it into a 3D cell. This explainer commits to numbers on both sides of that cliff: a cost ladder with multipliers, a tip-gauging station worked through to the micron, and the diameter below which robot guidance stops being reliable.

The short answer: presentation sets the price

Yes — the cost cliff is real, and how the bits arrive sets it. Bits fixtured end-up in a known nest need one 2D camera, one telecentric lens and a taught robot pose; the arm does nothing it could not do before. Bits the robot must find need a 3D sensor, hand-eye calibration, a gripper that tolerates pose error, and a retry loop for failed grasps. That stack — sensing, software, integration — is where the money goes, not the camera.

The index prices the vision-and-controls portion of the bill — hardware plus integration — against the fixtured baseline, and the whole cell tracks it roughly, because the arm is already on the floor. The top two rows are the planar-guidance case laid out in 2D vision pick and place: selecting a planar guidance system. The bottom row has a hard floor: below about 2 mm diameter it stops being achievable at all, for the reason in the small-bits section below.

Why fixtured, end-up presentation keeps the cost down

Four things make the fixtured cell quick to integrate:

• One nest fits every size. A V-groove rail or a magnetic plate holds any diameter in the range standing point-up, so metric and inch series ride through together with no tooling change and no recipe.
• Repeatability replaces guidance. A machined V-groove presents bits within about ±0.5 mm in X-Y and ±2° in tilt, so the arm replays one taught pose; vision only inspects, it never steers.
• The gripper is simple. A concentric magnetic or collet gripper closes on the shank 10–15 mm below the point. Roll angle is irrelevant and the point touches nothing, which matters when the point is the feature you gauge.
• The inspection hardware is stationary. The tip station bolts beside the nest, and the arm can present the held bit to the same lens again after placing.

Integration runs 1–2 weeks: one gauging routine, pass/fail logic, one robot pose per nest.

The 118° tip station, with the arithmetic

A drill point is a cone. Viewed from the side against a backlight, the cone projects as two straight flanks; you measure the included angle between them and the diameter where flank meets shank. Gauging cylinders this way is done with telecentric lenses, because magnification then does not shift when the bit's height or tilt wanders inside the nest tolerance — the trade is drawn in when you need a telecentric lens for gauging.

For a 118° point, tan(59°) = 1.664, so the visible point is 0.30 × D: 3.9 mm tall on a 13 mm bit, 0.15 mm on the smallest one. That ratio drives every choice below.

For the camera, take the MV-CS200-10UM: a 20 MP USB3 camera with a 1" IMX183 sensor, mono, C-mount. Rolling shutter is acceptable here because the nest indexes to a stop under the lens — keep the backlight on continuously and the exposure in the low milliseconds.

The IMX183 sensor lays out 5472 × 3648 pixels at 2.4 µm, a 13.1 × 8.8 mm imaging area. For the lens, the LS-TCL-2194 is a 0.5x C-mount telecentric for 1" sensors with a 110 mm working distance and a built-in coaxial light.

Now the arithmetic:

• Field of view: 13.1 mm / 0.5 = 26.2 mm wide, 8.8 mm / 0.5 = 17.6 mm high. Every diameter up to 13 mm plus its nest fits one frame.
• Object pixel: 2.4 µm / 0.5 = 4.8 µm. The smallest bit spans 500 / 4.8 ≈ 104 pixels, and its 0.15 mm point still covers 31 — enough for a presence check and a ±1.5° angle read. The 13 mm bit spans 2,708 pixels, its point 813.
• Accuracy: sub-pixel edge fitting holds edge repeatability near 0.2 pixel, about 1 µm here. After a one-time scale calibration against a pin gauge, plan on ±0.02 mm on diameter and point height — 2.5× inside a ±0.05 mm requirement — with angle repeatability of ±1–1.5° on the smallest tip and under ±0.5° from 6 mm up.

The pixel-on-the-object step generalizes to any camera-lens pair; calculate object space pixel size from magnification works it in both directions. The silhouette wants a collimated rather than diffuse backlight for knife-sharp edges — the comparison lives in collimated backlight or diffuse backlight for edge gauging. A full timing breakdown for a small-part gauging cycle is in measuring points on a 5 mm part in under one second.

Two upgrades cover special needs. If the angle tolerance on the smallest sizes tightens below ±1°, add a second, higher-magnification look. Use the LS-TCL-0811: an object-side telecentric at 1.36x with a 73 mm working distance for 2/3" sensors and a coaxial port. On a 2/3" sensor with 3.45 µm pixels that is 2.5 µm per object pixel over a 6.5 mm field, so the smallest point covers 60 pixels. Where the station must stand clear of the arm's envelope, the LS-TCL-1972 is a 1x C-mount telecentric for 2/3" sensors with a 440 mm working distance and a 177.5 mm body. At 1x a 2/3" sensor sees 8.8 mm across — the point region of any bit in the range.

The three lenses side by side:

All 3 share: Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

What each looser tier of presentation adds

Loose but lying flat, one layer, not touching. A top-down 2D camera now does the finding: X, Y and heading come from the silhouette, and Z is simply the belt surface. Cover a 60 × 45 mm tray region at a 300 mm working distance with a 1/1.7" sensor of 7.4 mm width: magnification 7.4 / 60 = 0.12, focal length 300 × 0.12 / (1 + 0.12) = 32 mm. The nearest standard focal lengths are 25 mm and 35 mm; take the 25 mm lens, which still covers the full 60 mm with margin.

The MV-CU020-90GC fits the tray role: a 2.5 MP GigE camera with a 1/1.7" sensor, global shutter, C-mount.

A 2.5 MP frame over that 60 mm field resolves about 29 µm per pixel: the smallest bit is 17 pixels, the 13 mm bit 444 — findable, not gaugeable, which is why the tip station stays telecentric. Light the tray evenly with ring lights; the LS-RNG-0305 is a standard ring light, 90 mm outer diameter, 44 mm inner, 24 V.

At a 100 mm standoff that ring flattens the field across the 60 mm width and tames the specular streak off ground flutes.

Piled in a bin or tote. A 2D camera cannot give Z here, so a 3D snapshot sensor enters. Budget: scan 0.5–2 s, pose solve 0.2–1 s, approach with clearance moves, and a retry loop — plan on 90–98% first-attempt grasp success, so a fixed share of parts gets scanned twice.

Fluted bits nest into one another, which stacks singulation on top of localization; allow a collision-aware gripper and presentation work before the scan. Integration grows from 1–2 weeks to 3–6, and the vision-and-controls bill lands at 4–8× the fixtured baseline. The sensor trade-offs are covered in 3D bin picking for flat parts: choosing a vision approach, with the caveat that flat parts are the easy case and cylinders are not; bin picking non-planar parts with 3D vision guidance goes deeper on pose quality. Presentation does most of the work in any 3D scan — layer height, spacing, finish — and configuring part presentation for robot 3D scanning works through the parameters.

The smallest-bit problem: where 3D guidance stops working

A snapshot 3D sensor resolves about 0.2–0.5 mm laterally at a 300–600 mm standoff. The smallest bit is one to two samples wide — noise, not a pose. A 1 mm bit is barely better; 2 mm is the practical floor for a graspable pose on a fluted cylinder, and entanglement below a few millimetres pushes the real limit higher. Commit to this split: robot-guided picking for 2 mm and above, fixture or tray feeding below it. The small half of the range — everything under 2 mm, including the whole 1/16 in to 9/64 in inch series — belongs in nests or an indexed tray.

The same small bit is easy in 2D: 17 pixels in the tray camera, 104 in the tip station. Thin steel cylinders under a robot share most of their handling problems with screws, and screw thread inspection camera selection for robot handling covers grippers and presentation for that neighbor case.

How to choose, case by case

• Bits arrive in nests, or an operator loads them: the fixtured tier. Buy one camera, one 0.5x telecentric lens, a backlight and a strobe-capable controller; teach one pose. Start at the configurator's lens step and browse the lens finder.
• Nests ride a belt whose position wanders: still 2D. Same hardware, plus a calibrated X-Y-angle correction output. No 3D anything.
• Bits lie loose in one layer: the tray tier. Add the top camera and ring light; keep the tip station as the robot-presented second step.
• Bits arrive piled and upstream will not change: the bin tier for 2 mm and above, with a feeder for the smaller half — and expect the 4–8× bill.

The rule underneath all four rows: if you can influence how the bits are presented, spend that influence before you spend it on sensing.

Checks before you order

• Fix the pass band. Set the point-angle limits at 118° ± 2° unless the drawing says tighter; production grinding typically holds that.
• Measure the nest. Within ±0.5 mm and ±2°, teach-and-replay works. Up to ±2 mm of drift, add the calibrated correction output instead of a 3D sensor.
• Do the blur sum. Blur = speed × exposure: at 100 mm/s keep any on-the-fly strobe at 20 µs for 2 µm of blur, under half of the 4.8 µm object pixel.
• Match image circle to sensor diagonal: a 1" sensor needs 15.9 mm of circle, an APS-H sensor about 27 mm. Most C-mount telecentrics sit in the 11–16 mm band, so pair that band with a 1" sensor, not a larger one.
• Plan both inspection poses now — the nest view before pick and the held-bit view after place — so the robot program and the trigger wiring land in the first integration week.

When the low-cost answer is the wrong answer

Fixturing loses in two cases. First, when the bits already arrive loose in totes and nobody upstream will change: a vibratory feeder for the whole mixed range is a project of its own, and the single-layer tray tier is usually the better middle — a scissor-lift tray, one layer, a 2D camera, still no 3D sensor. Second, when demand outruns an operator loading nests: a person keeps up to roughly one bit per second, and past that you need automated feeding, which is exactly where the 2–3× tier starts paying for itself. With bits under 1 kg and no size sorting required, neither case usually applies — the fixtured cell is the right answer.

Next step: run your numbers through the camera finder or start the configurator; to plan a cell across the tiers, send the part family and target rate via the contact page.