This is a little robot battle simulation, made with Unity Machine Learning Agents.

The motivation for this project was to experiment with using the camera's depth texture as visual input for reinforcement learning. I also wanted to create a quadruped agent with a rather small action/observation space, compared to the Crawler included in the ml-agents examples.

The project utilizes the now outdated version 0.5 of ml-agents. Please note that Agent.cs was changed to allow for a custom texture replacing the camera feed.
In order to run the trained models, you need to install the Unity TensorFlowSharp Plugin and add ENABLE_TENSORFLOW to Edit > Project Settings > Player > Configuration > Scripting Define Symbols in the Unity Editor. An archived copy of the plugin is available here.

Robot design by Lacomap (modified) via free3d.com