[simsla]

TL;DR: making industrial (e.g. manufacturing) robots easier to use, by improving sensing, planning, etc.

I suspect that Dr. Chelsea Finn's work in meta-learning (affiliated with Stanford and GBrain, when I saw it last year) might play a big part here, which is e.a. about generalisation of RL policies to out of domain tasks. (E.g. similar task, but slightly different tools, slightly different task, etc.)

Learning IRL (cameras and actuators) reinforcement learning policies is a huge time sink, so generalisation is a hugely important task. Related solutions can be found in simulation->real generalisation, also an active topic of research.