We were fascinated by the concept of teaching an AI to interact with the physical world. We wanted to learn Unity and test our skills in Reinforcement Learning at the same time.
What it does
It proves that it is possible use Unity as a simulation environment to quickly create physics focused tasks and test reinforcement algorithms on those tasks.
Challenges we ran into
All of them. Specifically, reinforcement learning was slow to run on our old laptops, and the structure of the project did not make it easy to develop on two computers simultaneously.
Accomplishments that we're proud of
We are proud that were able to learn ML-Agents and stay focused during almost the entire hackathon.
What we learned
We learned how to create arbitrary environments in Unity that can interact with reinforcement learning algorithms.
What's next for N-tuple Pendulum Dampening with ML-Agents
Porting our environment to an open-source "gym" so that researchers and enthusiasts can test out their RL ideas. We also want to do some further hyperparameter and reward tuning to create a strong baseline for others to compare against.