I saw these folks talk at a conference last year and demo some of this work. There is a strong push towards asking computers to do tasks that humans and animals do in psychology and neuroscience experiments: navigating a maze, finding a platform submerged in water, etc. Many of these tasks require memory (see also: playing Montezuma's Revenge). Part of the goal is to reach parity with humans on certain tasks. Another part is to help understand how cognition works by building it from scratch and cross-referencing results with human/animal results.

By using Unity, they can create a virtual environment that mimics the real physical experiments. I think this is good, as one big 'cheat' in many AI projects is representing the 'state' of the system in a way that implicitly captures expert domain knowledge - i.e. about what parts of the environment are significant with regards to performing the task. Letting the agent 'experience' a virtual world in a 'raw' form allows it to figure out what's salient - what it needs to remember - which animals also have to do in experiments.

Plus, Unity videos demo really well.

Traditional machine learning simulators can be hard to work with. Meanwhile unity has been pushing to support machine learning development, and their world construction tools are great. It’s a lot easier to build a little car and a suitable world in unity than it is in pybullet, for example.

They did a deal a year ago: https://blogs.unity3d.com/es/2018/09/26/unity-and-deepmind-p...

same here, although i know unity has been staffing up their ml team eg arthur juliani doing good work at unity. more generally speaking, i think there is a clear need for reliable, robust simulators in ml. (openai released a few interesting environments a few years back but adoption seems to have plateaued, from my perspective.)