Why Vision Guidance Is Essential for Flexible Automation
Classic robot programming assumes exactly known, repeatable positions. As soon as parts sit unsorted in a bin, arrive irregularly on a conveyor, or come from different batches with slightly varying tolerances, pure teach-in programming is no longer sufficient. Vision-guided picking solves this by having the robot recognize an object's actual position and orientation in real time and adjust its path accordingly.
2D Vision: The Economical Standard for Defined Planes
2D camera systems detect an object's position and rotation on a known plane - for example on a conveyor or a flat pallet. They are less expensive, easier to calibrate, and fast enough for high takt rates. They reach their limits once objects vary in height, are stacked or overlapping, or when the grip position depends on the object's orientation in space (not just in the plane).
3D Vision: Necessary for Unstructured Scenes
3D systems (stereo cameras, time-of-flight cameras, structured light, or laser scanning) capture a complete point cloud of the scene. They are a fundamental requirement for classic bin picking, where parts lie unordered at varying heights and orientations inside a container. Computational load is significantly higher than with 2D systems, but modern embedded vision processors and optimized point-cloud algorithms have reduced cycle times considerably in recent years.
Selection Criteria in Practice
1. Degrees of Freedom Required in the Gripping Scenario
If the robot only needs to grip in X/Y and one rotation, 2D is generally sufficient. Once tilt angles, height variation, or object overlap come into play, 3D becomes necessary.
2. Required Cycle Time
3D bin-picking systems typically need 0.5–2 seconds to capture and calculate the optimal grip point. At very high takt rates, pre-sorting with simpler 2D detection or mechanical singulation can be more economical than pure 3D bin picking.
3. Material Properties
Highly reflective metal parts, transparent plastics, or deep black surfaces present a challenge for many 3D sensor technologies. Structured light often performs worse on shiny metal parts than laser-line triangulation. Testing with the actual part before selecting a system is essential.
Common Integration Mistakes
The most common mistake is insufficient hand-eye calibration (eye-in-hand or eye-to-hand). If calibration is performed with a different tool or gripper than the one used in production, systematic offset errors emerge that only surface as collisions or inaccurate grips. Equally critical is inadequate lighting: changing ambient light, reflections from factory windows, or shadows cast by the equipment itself lead to inconsistent detection rates. Controlled lighting decoupled from ambient conditions is mandatory in production systems, not optional.
Integration with KUKA, ABB and FANUC
All three manufacturers offer native vision interfaces: KUKA.VisionTech, ABB Integrated Vision, and FANUC iRVision. The choice between a manufacturer's own solution and an external vision system (e.g. Cognex, Keyence, Zivid) depends on integration effort and desired flexibility. Manufacturer-native systems are generally easier to commission, while external systems often offer more capability for complex 3D scenes.
Conclusion
Vision-guided picking is now standard practice for any application with variable part positions. The choice between 2D and 3D should not be based on availability but on the actual degrees of freedom the application requires. Clean calibration with the final tool and controlled lighting determine success or constant rework in practice.