Short answer: choose 2D vision when the part's pose is controlled — presented flat, at a known height, in a limited range of orientations — and choose 3D vision when the part's height, tilt or stacking varies and the robot must measure it before moving. A 2D camera answers "where is it in the plane, and rotated how?"; a 3D system answers that plus "how high, and leaning which way?". Everything else — cost, speed, lighting, software — follows from whether your application actually asks the second set of questions.
Where 2D vision is the right tool
2D machine vision is the mature, fast and economical default for a large share of robot guidance. It excels when the world is arranged for it: components on a conveyor or tray for pick-and-place, labels and codes to read at a fixed distance, parts located against a backlight silhouette, surface inspection where the defect is visible as contrast. Vendors such as Cognex built the category on these applications, and camera makers such as Basler supply the high-speed, high-resolution sensors behind them. Cycle times are short because a single exposure is all the data there is; lighting is a solved discipline of domes, bars and backlights. If a fixture, feeder or tray can present your part consistently, 2D is usually the fastest and cheapest correct answer — and money saved on vision is better spent on the presentation hardware.
Where only depth will do
3D vision earns its price when pose variation is the problem itself. The canonical case is bin picking: parts arrive jumbled in a container, and the robot must find one part's full position and orientation, plan a grip that avoids its neighbours, and extract it. Pallet work is the same shape of problem — cartons stacked to varying heights, layers that shift, stretch film that confuses flat imaging. Add machine tending of raw castings, volume measurement, and locating parts on curved or stacked surfaces, and the pattern is consistent: whenever the robot's safe approach depends on height and tilt, a flat image is not enough. The GlobleMatch directory lists 54 companies under machine vision and 3D sensing (as of 10 November 2026), most of them built around exactly these applications.
The two side by side
| 2D vision | 3D vision | |
|---|---|---|
| What it measures | Position and rotation in a plane; presence, codes, surface contrast | Full pose: position, rotation, height and tilt; volume and shape |
| Part presentation | Must be controlled (tray, feeder, conveyor, fixture) | Tolerates random presentation (bins, pallets, stacks) |
| Speed | Fastest; single exposure, simple processing | Heavier data (point clouds); modern systems are fast, but total cycle time must be tested |
| Lighting sensitivity | High; lighting design is half the project | Less dependent on ambient light for depth, but reflective and transparent surfaces are hard |
| Cost shape | Lower sensor cost; budget goes to lighting and presentation | Higher sensor and software cost; can remove presentation hardware |
| Typical failure | Unexpected pose variation breaks the setup | Shiny, black or transparent parts produce poor point clouds |
The hybrid most factories end up with
Real cells often combine both. A common pattern: a 3D camera locates a part in a bin or on a pallet and guides the rough pick; a 2D camera on the wrist or at a re-grip station then refines orientation before placement or assembly. Another: 2D handles the fast main task (reading, locating, inspecting) while a compact depth sensor watches a single dimension that matters — fill height, stack height, presence of a raised edge. Combining is not indecision; it matches each sensor to the cheapest question it can answer reliably. The gripper choice interacts with this decision — a tolerant gripper can forgive small pose errors that would otherwise force a vision upgrade, which is why vision and gripper selection should be specified together.
Five questions that settle the choice
- How much does part pose vary when the robot first sees it? If fixtures bound the variation, stay 2D.
- Does the robot need height before it moves? If yes, you need depth from somewhere — sensor or mechanics.
- What are the surfaces? Mirror-finish metal, black rubber and clear film are the hard cases for 3D; demand a trial on your parts.
- What cycle time must the whole loop hold? Measure capture-plus-processing on your part, not the brochure frame rate.
- Can presentation hardware remove the variation more cheaply? A feeder is often the best "vision upgrade" available.
Suppliers across both camps are listed side by side in the machine vision section of the directory, so you can shortlist 2D and 3D vendors against the same application description before committing to either route.
Frequently asked questions
Can a 2D camera guide a robot to pick parts?
Yes, when the parts are presented in a controlled way — flat on a conveyor or tray, at a known height, in a limited range of orientations. Most high-speed pick-and-place in electronics and packaging works exactly this way. The camera finds position and rotation in the plane; the robot's approach height is fixed by the setup.
When is 3D vision unavoidable for a robot?
When the part's pose varies in all dimensions: parts jumbled in a bin, boxes stacked at varying heights on a pallet, or workpieces that must be located on a curved or uneven surface. If the robot needs to know height and tilt before it moves, it needs depth data — that is the dividing line.
Is 3D vision slower than 2D?
Historically yes: capturing and processing a point cloud takes longer than a single 2D exposure, and the gap matters at high cycle rates. Modern 3D cameras have narrowed it considerably, but for the fastest flat-part applications 2D remains the default. The honest comparison is total cycle time on your part, measured in a trial — not the camera's headline frame rate.
What is the most common vision mistake in robot projects?
Buying the sensor before defining the presentation of the part. Many projects that "need 3D" actually need a feeder, a tray or a fixture that presents parts consistently — after which a simpler, faster 2D system does the job. Fix the presentation first, then buy the vision the remaining variation requires.