aws-samples / amazon-sagemaker-unity-game-simulation

unity-roller-ball-simulation

This is a sample unity project for Unity Game Simulation with ML-Agents running on Amazon SageMaker RL.

Preparation

Download this repo

First, download a zip file from the repo and open it.

You will find a folder called unity-roller-ball-simulation-master.

Unity

Start Unity Hub. In Projects window, press Add and select RollerBall folder in the folder described above.

You may need to download the version 2019.3.14.f1 version of Unity with Linux support module.

This is how the sample game, RollerBall looks like.

There are three important scripts, ObstacleManager.cs, RollerAgent.cs, and RollerBallConfig.json.

RollerBallConfig.json is found in Resources/Congig/.

Build the project for Linux platform

Open Player Settings (menu: Edit > Project Settings > Player).

• Under Resolution and Presentation, ensure that Run in Background is checked.

Open the Build Settings window (menu:File > Build Settings).

• Select Linux as Target Platform
• (Optional) Development Build is checked.

Press Build And Run

We name a collection of built files as RollerBall_build_000.

Here is how the built files look like in the RollerBall folder.

Copy the flowing files and folder and upload them to S3.

• LinuxPlayer_s.debug
• RollerBall_build_000_Data/
• RollerBall_build_000.x86_64
• UnityPlayer_s.debug
• UnityPlayer.so

Please take note of S3 URI. You are going to need that for rl_unity_cloud_simulation_sample.ipynb

Amazon SageMaker RL

Amazon SageMaker RL on SageMaker notebook

Start Amazon SageMaker and create Notebook Instance. (menu:Notebook > Notebook instance > Create notebook instance).

• t3.xlarge instance is adequate.

Click Open Jupyter and select SageMaker Examples

SageMaker provides dozens of sample notebooks. You can find a samle notebook named rl_unity_ray under Reinforcement Learning tab.

Select rl_unity_ray and press Use botton. A new directry will be created.

/home/ec2-user/SageMaker/rl_unity_ray_YYYY-MM-DD

Go to rl_unity_ray_YYYY-MM-DD

Upload the following files of this repo to the directory above.

• rl_unity_cloud_simulation_sample.ipynb
• Dockerfile
• entrypoint.sh

Replace the following files of the directory above with those in this repo.

• src/evaluate-unity.py
• src/train-unity.py

Proceed rl_unity_cloud_simulation_sample.ipynb

You are ready to start the notebook.