We address the problem of autonomously learning to perform tasks. This is the fundamental challenge in developing genuinely autonomous systems across fields, from robotics and autonomous vehicles to customer services, employable outside the narrowly defined tasks the current hand-crafted solutions require. The challenging bottleneck is that autonomous learning necessarily involves exploration, which in the real world is costly and unsafe. This is the reason why the impressive demonstrations of recent deep learning powered reinforcement learning are not yet available in the real world. B-REAL aims to bridge this reality gap between autonomous learning in real and simulated worlds.
The core idea is to learn an artificial surrogate model, in which both performance and safety can be checked before real-world action. What is new in this project is that now we believe we are in the position of setting up the currently disjoint modelling methods required for coping with the reality gap into a complete next-generation autonomous system. This requires efficient ways of learning accurate simulators, matching them with the real world, and including safety constraints. Safe and effective operation requires taking into account the uncertainty of the simulation outcomes, which is now possible even for the necessarily flexible simulators due to our recent, award-winning breakthroughs in deep probabilistic modelling.