How to train ABB robot by deep learning models?

Hi,
You wouldn't train the robot for this. The controller doesn't have the computational power required to train any significant deep learning model. And it does not have any machine learning libraries that would help you with it.

Instead, put the deep learning part on to a dedicated edge computing device like the Nvidia Jetson SBC's that is connected to the camera(s) and does the inference part of object detection/ tracking or whatever you require since you can run multiple networks in parallel it is possible to extract all the computations you could want even if they wouldn't be suitable to train with one and the same model. By outsourcing the machine learning to a dedicated edge device you unload this computationally intense task from the controller. You will be able to leverage from all the powerful frameworks out there like PyTorch and TensorFlow when building the model it could also be interesting for you to take a look at intel's OpenVino toolkit for running the inference part when the training is done.

The training part is the computational heavy part of deep learning so if you're about to train the network from scratch you either want a really powerful computer with loads of memory to speed up the process or run the training as a cloud instance.

A way of speeding up the prototyping stage is to try simple transfer learning on one of the pre-trained models out there that already been trained on a dataset like ImageNet which contains about 13million pictures divided into 1000 classes. By leveraging transfer-learning you keep the knowledge from that model and just retrain the last layers to fit your custom object detector.

Hope this was helpful to you.