A camera can locate a part, confirm its presence or inspect a defect. These tasks require different criteria. Before choosing a sensor, define the decision the robot must make and the part variation that mechanical positioning does not control.
Define the required measurement
For picking, describe the target point, allowed orientations, accessible face and placement tolerance. For inspection, define the defects and the pass, fail or undetermined decision. A presence check does not establish the overall quality of the part.
Check whether mechanical guides, a fixture or more controlled feeding would be sufficient. Vision becomes an option to test when the remaining variation requires measurement. Selection then depends on the actual parts and how they arrive.
Choose the dimensions that matter
A 2D measurement may suit parts presented in a controlled plane. Cognex In-Sight Explorer 6.3.2 VGR documentation distinguishes planar X, Y and angle from the other coordinates in its interface. This description applies to that software and does not establish direct compatibility with an ESTUN controller.
Variable height, stacks or orientation outside the plane may require depth measurement. Cognex documents point-cloud tools and 2D projections in In-Sight Vision Suite 25.2. Occlusion, surfaces and tool access still need testing on project samples.
Connect camera, part and robot coordinates
The project file should identify camera, robot, tool and work-plane coordinate systems. Calibration relates these frames. The Cognex procedure consulted illustrates hand-eye calibration for a specific Universal Robots configuration. It is a method reference, without transferring its commands to another robot.
Image resolution alone does not determine final picking error. Measure that error at several positions within the working area, using the intended mounting, lighting and load. Define which events require calibration checks.
Prepare the trial campaign
Prepare samples covering expected colours, surface conditions, orientations and positional variation. Include difficult parts, partial occlusion and no-part conditions. Separate tuning samples from those reserved for verification.
For each trial, retain the image or measurement, expected decision, actual decision and observed pick. Count false accepts, false rejects and undetermined results separately. Agree acceptance thresholds before the campaign.
Handle missing or stale results
The program must associate each result with the correct part and request. Define acknowledgements, timeouts and pose validity conditions. Stale results, interrupted communication or a part moving after capture require explicit behaviour.
Recovery should specify whether a new acquisition is required, who may remove a part and how the cycle returns to a known state. Include these cases in trials with the cell safeguards active.
Accept the complete chain
Acceptance concerns the placed part under agreed conditions. It covers acquisition, pose calculation, motion, gripping and failure handling. A good camera result can still lead to a poor pick if the tool or coordinate frame is incorrect.
Hand over versions of the relevant programs, settings, recipes and calibration files. Record tested formats and observed limits. This file should support assessment of a new product or lighting change.
Common questions
Points to clarify.
Is a 3D camera always necessary?
No. Presentation in a controlled plane may allow 2D measurement. Variable height and orientation need assessment on actual parts.
Is resolution enough to select a system?
No. Lighting, calibration, surfaces and gripping influence the result. Acceptance measures the complete chain.
What should we prepare for a vision study?
Representative parts, expected variation, target defects, tolerances and installation conditions.
