This project won the 3rd Prize of 2019 Global VR Hackathon held by VRCORE.
The idea of this project is to create a better VR interaction experience with objects. Instead of hardcoding the behaviors of the objects, we collect player data and utilize Unity ML-Agents Toolkit to train a model that improves how the objects respond to players by reinforcement learning.
In the beginning, we trained the agent to fly without losing control. Once the agent could keep itself balanced, we started to train it with more inputs such as the position, the rotation, and the velocity of the weapon. After six hours of training, the agent behaved like it knew where the weapon was coming from and could dodge the attack a little.