Transformers for snap-fit detection : Deep learning for teaching robots to click components into other components

Sammanfattning

If living standards are to continue to improve then more of our manufacturing has to be automated. A difficult problem to solve for automation has been joining two plastic components by pushing one into another until they click into place. This is due to variations in components and and because of this, the task is done manually today. With the resurgence of artificial intelligence a lot of previously difficult to automate tasks have become viable for automation. While some work have been done in the field a lot of work remains in regards to generalization. This master thesis aims to contribute to the field of artificial intelligence by applying a transformer model based on a vision transformer as well as to recreate the results of another paper by using a convolutional neural network on the same problem. The transformer based model is tested in two variants, one more closely resembling the vision transformer and one simplified by removing the trainable the class token from the input. The experiments achieved a test time accuracy of 100% for the convolutional neural network based model as well as both variants of the transformer model. A detailed comparison between approaches revealed that the transformer based model more reliably achieved the 100% test time accuracy, with the simplified variant being even more consistent. The experiments also revealed that performance would vary depending on the percentage of data reserved for training with more training data not being beneficial for test time accuracy past a certain point.