Answer from AI Robotics Australasia. .
| Feature | 2D vision | 3D vision |
|---|---|---|
| What it measures | Position and features on a plane | Shape and pose, including depth |
| Typical sensor | Industrial camera and lighting | Structured light, stereo, or time of flight |
| Good tasks | Presence, labels, flat pick from a conveyor, many surface checks | Random bin picking, depalletising unknown stacks, unfixtured pose |
| What breaks it | Depth variation the camera cannot see | Shiny, transparent, or very thin parts that return a bad cloud |
| Cost and complexity | Lower, if the lighting is actually controlled | Higher, and the grasp plan has to use the cloud |
Stay with 2D when the world is flat
A well-lit camera finds a part on a conveyor, reads a code, or checks that a feature is present. Adding a third dimension does not make that task more reliable. It adds a sensor, a calibration, and a new way to fail.
Lighting is the work. A 2D system with a backlight and a repeatable presentation will beat a 3D system pointed at a glossy part in ambient factory light. Specify the presentation before you specify the sensor.
Move to 3D when the part is not presented
Random bin picking fails in 2D because the part you want may be tilted, buried, or one layer down. The robot needs a surface it can approach and a grasp that does not collide with the bin or the neighbouring part. That is a point cloud problem, plus a grasp planner, plus a rule for when to shake, re-scan, or skip.
Depalletising a known pattern can be simpler. If every layer is the same height, a 2D find plus a layer counter is sometimes enough. If layers are mixed or the stack is slumped, you are back to 3D.
The sensor is not the cell
Vision projects stall on reflections, clear packaging, black rubber, and parts that hook together. Name those parts in the trial. A demonstration on a matte sample is not evidence your real bin will run.
We trial the actual part. 2D, 3D, or a combination is chosen from the pick, not from a preference for one sensor family.
Related questions
Do I need 3D vision for bin picking?
For a jumbled bin, yes. The robot has to know the height and tilt of the part it is about to grasp. A structured presentation in a tray can sometimes be done in 2D.
Is 3D vision better than 2D vision?
Only when depth is part of the problem. For labels, presence, and flat picks, 2D is usually more reliable and easier to keep running.
Why does robot vision fail on shiny parts?
Specular highlights and transparent surfaces confuse both cameras and many 3D sensors. The fix is part of the trial: polarisation, a different sensor, a coating, or a presentation that hides the reflection. It is not something to discover after installation.
Can 2D vision depalletise?
Yes, when the pattern and the layer height are known. Mixed or slumped stacks need 3D, or a human still in the loop.
Related systems
- AI Bin PickingRandom bin picking with 3D vision and spatial AI. Handles mixed SKUs, irregular shapes, and cluttered bins without fixed fixtures.
- AI Quality InspectionInline visual inspection with deep learning defect detection. Catch surface flaws, dimensional errors, and assembly defects at line speed.