Have a look at our Google Drive link in the description below.
Please have a look at the following git repositories: 1) https://github.com/jacqueselliott/unityRL 2) https://github.com/96imranahmed/asyncRL 3) https://github.com/jacqueselliott/SpatialRL
Inspiration
Deep Learning is powerful - Deep RL can solve real complex stochastic problems and we wanted to show a proof of concept that Deep RL can be used with Unity
What it does
Have a look at the G-Drive. There are a lot of moving cogs (especially with Spatial OS)!
How we built it
Come around to our table and we'll be happy to tell you!
Challenges we ran into
SpatialOS - Spent a lot of time trying to get C++ workers to work. Made great headway but couldn't quite make it all compile due to errors with C++ Tensorflow API. Code works in part although integration has been very difficult due to the novel nature of all the software packages we were using
Accomplishments that we're proud of
Managed to create a complex Deep-RL architecture that trained successfully on a Unity simulation. We also got the (notoriously difficult) Tensorflow C++ API to work!
What we learned
Deep RL + Spatial OS + Unity is a powerful combination that can be used to train complex AIs in a scalable manner in the future.
What's next for SpatialRL
More research. More layers. (Much) More hyperparameter tuning!
But more seriously, we want to dive in more into multi-agent problems with more complex simulations because we believe there is a lot of scope in this area!
Log in or sign up for Devpost to join the conversation.