I don't think its a training session. Current AI models are pre-trained before deployment for inference. After the model is trained, they load it into the robots computer, and it runs inference with that model. You can't train the model again because you don't have enough memory on the robot, but also even if you did its slow and consumes energy. You could have it train in some server but then every new skill would require you to pay the equivalent price for renting a bunch of GPUs for many hours.

What they can do is, for everyone, have a base model, and then improve it over time. Then, with software updates they can improve the set of skills the robot can handle out of the box.

But this is the problem with current AI systems, without a continuous learning capability, you're always limited to the "default skills". As soon as you have something out of the box for the robot to do, you end up needing Indians to learn it.

All of AI is flawed in this way. LLMs for instance have almost no continuous learning capability, that is why we don't have AGI yet. They can't learn new skills. Therefore, they can't adapt to new jobs they have not seen during training. They can't even play pokemon properly or any complex game for that matter, because games involve learning new skills during gameplay.