Teaching a round object with the Robotiq camera

Sebastien 2016-09-15

@Etienne_Samson
I am teaching an object that is round as seen from the top. Like a washer for example. When I am in the TEACH OBJECT screen and I take the required 5 snapshots, I run into issues. In this interface you are asked to first take a picture of the background, then add the part and take another snapshot. Rotate the part 90 degrees and take another one and so on. However, since the part in round, it does not matter if you turn it to 90 degrees, it is still in the same orientation. So for the first picture of the object it is fine, but for the others sometime I run into an error saying 'Failed to refine model'. Even though sometimes I did not touch the part at all between snapshots. What does this error mean and how can I get around this?

Etienne_Samson 2016-09-15

@Sebastien I have seen this issue when the part is reflective, is it ?

Here is the explanation: The reflection on your part is seen as a feature by the system, thus your part is not symmetric with that reflection. If you don't turn the part, it fails. If you turn it, then the reflection does not follow, feature is not recognize, thus it fails.

What you should do: If your part is reflective (metal) find a spot on your workspace where there is less reflection. Eliminate reflections sources (ambient lights). In extreme cases, go to the source and eliminate reflections from the part, paint your part with a mat paint. Remember that the system sees the geometry, so it doesn't mater if you teach a black part and then pick a raw metal part. With that you should be able to teach it.

One last trick, sometimes going the opposite works too, put full lights on the part, if it is reflective, with enough lights it will be seen as a uniform white part. You should then also be able to teach it.

Catherine_Bernier 2016-09-15

@Sebastien and @Etienne, I have been discussing this with @Vincent_Paquin and one trick would be to tilt the robot a bit. When viewing a round object directly from above, you see a circle. However when you tilt a bit the camera, the circle changes to an ellipse. This allows the vision system to snap a model more easily because it is always looking for an object’s orientation and this is harder to find with circles than it is with ellipses.

Applying this concept to the post Square Object Using the vision system tilting the camera's snapshot position a bit will make the square rectangular and therefore limiting 1 degree of freedom when finding the object with the camera.

However, we have to be careful not to tilt too much because this will also have some impacts on location accuracy.

@Vincent_Paquin, any tips to add?

Etienne_Samson 2016-09-15

@Catherine_Bernier this trick will work too but only for planar objects, perspective effect (the circle that becomes an ellipse) will not be compensated enough if the object has any height. So if you have a flat part, go for it, otherwise don't do this.

Vincent_Paquin 2016-09-15

Tilting the robot a bit will help with reflexions coming from the camera's lights, but the perspective effect is cancelled by the calibration. So when the pattern matching is computed, a circle is still a circle even under perspective. Same thing for a square. What really happens when using perspective is that you will add features from the object sides that are usually not visible directly from above and this can help in some specific cases.

Note that the more angle you give, the more flat your objects need to be.