AI Snow Walker: A game about knocking over an AI that taught itself to walk

I made this game as a present to my colleagues on the MARS Next Generation Technologies Team .

Playable in-browser, this simple game lets you throw snowballs at an AI and test it’s balance:

Some fun facts:

  • This AI taught itself to walk using Deep Reinforcement Learning

  • It is made up of rigid body physics objects that are linked in a way to mimic the human body

  • I optimized the model for reaching its destination AND taking snowball hits!

  • I used Unity's ML agents Machine Learning package to train the AI

  • I used this example project from Unity and retrained the model from scratch with a hybrid curriculum based learning approach.

  • I used uneven terrain to force the model to generalize better than Unity's model

    • Ray Perception Sensors were used to give the model "Lidar" vision of the ground in front of it

Technology Used:

  • Unity Game Engine 🎮

  • Pytorch🔥

  • Unity Machine Learning Agents 🤖

  • Tensorboard 📈

  • Deep Reinforcement Learning 🧠

  • Love for my colleagues

Previous
Previous

AI Classifier for Healthy Coral Reefs: An AI Hackathon Challenge

Next
Next

Predicting Pet Life Expectancy: An AI Hackathon Challenge