Key takeaways
- Use 2D vision when parts sit on one plane with predictable height and you only need X, Y, and in-plane rotation.
- Move to true 3D when parts stack, overlap, tilt, or arrive in bulk with six-degree-of-freedom pose uncertainty.
- Reflective, transparent, or matte-variable finishes often force extra lighting or depth sensors, not more 2D resolution alone.
- Match vision processing time to cobot cycle time before you buy a faster arm or gripper.
- Start with the simplest vision that meets the pick proof, then add depth only when field data shows missed picks.
When does 2D vision win on a cobot line?
For high-mix cobot picking, two-dimensional vision is the default when every part shares a flat presentation surface and height stays within a few millimeters. The camera reports position in X and Y and rotation around Z. That is enough for tray-fed inserts, flat blanks on a conveyor, or kits where a fixture sets the pick height.
Two-dimensional systems are mature, fast, and comparatively inexpensive. Field guides for semi-structured picking often cite image acquisition and pattern matching well under a few hundred milliseconds when contrast is good. That headroom matters on collaborative cells where the arm itself is the slower element.
The tradeoff is depth blindness. If a tote contains layers, or a part can tip on its side, a flat image cannot tell the gripper how high to approach or which face is up. High mix then becomes a lighting and fixturing problem instead of a software problem, which is rarely cheaper.
What changes when you add depth sensing?
Three-dimensional vision reconstructs the workspace with height and surface angle. Structured light, stereo pairs, and time-of-flight sensors each trade speed, resolution, and sensitivity to glare. The integrator choice is less about the sensor badge and more about whether you need a dense point cloud to segment overlapping parts.
Depth pays off when presentation is messy: bulk bins, mixed returnable totes, or depalletizing where layer height drifts. The robot needs six-degree-of-freedom pose estimates to clear neighbors on the way in. Without depth, operators end up re-staging parts between shifts, which erases the labor savings you bought the cobot to capture.
Depth also costs more than hardware. Calibration ties the camera frame to the arm flange, gripper geometry, and workcell fixtures. Change any one of those without re-teaching, and pick quality drifts. Budget engineering time alongside the sensor invoice.
How do reflectivity and part finish steer the decision?
High-mix lines rarely fail because the arm is too slow. They fail because the scene violates the assumptions baked into the vision recipe. Polished aluminum, clear packaging, and oily stampings scatter light in ways that flatten a two-dimensional image into ambiguous blobs.
Depth sensors have their own glare modes. Structured light can wash out on mirror-like curves. Time-of-flight can lose fidelity on very fine edges. The fix is usually environmental: polarizing filters, matte carriers, controlled backlighting, or mechanical singulation before the camera sees the part.
Texture and color cues still help in high mix. A two-dimensional pass can identify a label, pad print, or molded feature while a depth pass confirms height and tilt. Many cells run a lightweight 2D identification step plus a 3D pose step only when the SKU library demands it.
Where does 2.5D fit between flat imaging and full 3D?
Between classic 2D and full point-cloud picking sits a practical middle ground: height-aware picking on a mostly flat bed. Parts may vary in stack height but still present a stable top face. Sorting washers on a tray, handling blister packs, or lifting lids from a flat conveyor often land here.
Two-and-a-half-dimensional setups add a height map or laser line without the full segmentation load of random bin picking. Cycle times usually sit between structured fixture picks and heavy 3D bin cycles. If your mix includes both flat kits and occasional jumbled totes, you may deploy 2.5D on the main line and reserve true 3D for a secondary debulk station.
How should you align vision speed with cobot cycle time?
Collaborative picking lives or dies on total cycle time, not camera megapixels. A vision pipeline that adds two seconds on every pick may be fine for machine tending with twenty-second cycles and unacceptable for a packaging line targeting eight seconds.
Tech industry guidance for collaborative vision stresses matching processing latency to the arm program. Capture, decode, robot communication, and gripper actuation all consume the same clock. Profile the worst SKU, not the easy one, before you freeze hardware.
Random depth picking often runs longer per cycle than structured 2D picks because segmentation, collision checking, and occasional re-grasps stack on top of motion. Planning for that gap up front prevents a pilot that looks brilliant on one part number and collapses when the mix widens.
What does high-mix picking cost in commissioning time?
Fixture-based 2D teaching can finish in days when lighting is stable and SKUs share a common envelope. Each new part still needs a trained model or pattern, so high mix pushes you toward flexible tools: CAD-based alignment, feature templates, or teach-from-example workflows.
Full 3D bin picking commonly spans weeks of tuning because gripper reach, bin geometry, and point-cloud noise interact. Simulation helps, but production pallets, dunnage, and worn bins show up only on the floor. Pilot with the dirtiest tote you allow in receiving, not a clean demo bin.
Service Robot Co. often stages these projects as phased cobot rental or month to month lease pilots so plants can prove pick rates on real mix before they standardize cameras across cells. A vendor-neutral integrator can pair the arm, gripper, and vision stack that fits the envelope instead of forcing one sensor brand everywhere.
How do you decide without overbuying sensors?
List the top ten SKUs by volume and the ten worst outliers by shape or finish. If nine of ten picks need only planar pose, buy 2D and invest in presentation. If three or more outliers arrive jumbled, model the cost of manual pre-sort versus 3D upfront.
Measure miss-pick modes on a shadow line: height errors, rotation errors, collisions with bin walls, or false positives on similar parts. The failure label tells you whether you need depth, better lighting, or simply a different gripper.
According to the International Federation of Robotics, global factory robot installations reached about 542,000 units in 2024, with electronics and automotive each near one quarter of demand. Vision-heavy cobot cells ride that same capital cycle: plants add sensing when labor mix and SKU count make fixed automation brittle.
- Document minimum and maximum part height in the pick zone.
- Record worst-case glare and shadow across a full production day.
- Time vision plus motion for the slowest acceptable SKU.
- Define who owns re-teaching when engineering changes a mold or label.
What should maintenance own after go-live?
High-mix vision fails quietly. Pick rates slip half a percent per week until someone notices scrap at the next station. Build weekly checks for lens contamination, lighting drift, and gripper wear. Depth sensors also need periodic recalibration after hard collisions or tool changes.
Keep golden images from acceptance testing. When a new SKU launches, compare its first picks against that baseline instead of tuning live on the production floor. Remote triage and on-site dispatch matter because a cobot cell without vision is just a paused arm blocking an aisle.
Treat vision parameters like any other production recipe: versioned, audited, and rolled back when a change misbehaves. The goal is not the fanciest point cloud. It is reliable picks across the mix you actually ship.
