I'm not sure what the gloves add to this scenario. The point of haptic gloves is to give the sensation of virtual objects. It sounds like this idea would work much better by simply pointing an Ultraleap camera at a human doing the task. You want hand tracking data from a human doing a real thing. Haptic feedback isn't helpful here.

Besides, I don't think that training a neural network on hand motions is actually that useful. What you're talking about is a kinematics problem that is totally tangential. You'd want the robot to understand how to manipulate objects, not hands, and not a human body. Watching how a human moves doesn't tell you how a robot moves. The neural network would need to learn how to manipulate arbitrary objects with a robot body. Capturing human motion doesn't help much.

> You want hand tracking data from a human doing a real thing.

You want pressure and force information too, for helping train the robot to be dextrous.