Supertuxkart ice hockey code#
Supertuxkart ice hockey simulator#

In total, we collected around 10,000 labeled images for use in training our model. For points outside of our image, we experimented with leaving the values as were, clipping them to the nearest valid x value, and setting them to 0. We repeated this process for several games, generating data from both team karts perspective and from both starting sides of the field. Using functions provided in previous assignments, we were able to both program the kart to follow the puck, and output the (x,y) screen coordinates the puck was mapped to. In order to generate labels for our training data, the pystk state instance was fed to each player kart instance, allowing them access to the true puck location. We used a setup with 2 (conv2d, batchnorm, relu) blocks per layer with 4 layers used. Our vision based approach started with a similar FCN model to the one that was submitted for HW6. In the end, we decided to focus our efforts on the vision and controller-based approach for our final submission. We felt the imitation approach was promising, but we had no way of estimating how much time investment it would take for it to become as performant as the vision and controller technique. Around this time, the vision and controller-based approach became performant: we were scoring one goal against the AI 1v1 about 50% of the time and beating the AI about 10% of the time. When the game was executed with 2v2, the AI seemed to break and behave erratically. However, for an unresolved reason, the AI only behaved as a quality oracle when we ran the game 1v1.

Supertuxkart ice hockey code#

In order to accomplish this, we modified some code from the Pystk code base and were able to collect labeled data with the AI playing. We did not believe manually playing the game could cover every state the kart could find itself in a reasonable amount of time. Ideally, we could leverage the game’s AI (set to difficulty level 2, the strongest) to collect a massive amount of training data. In the end, we had around 75 possible “classes” that our classifier could choose from, each corresponding to a different combination of actions our kart could take.Ī critical component of the imitation learning approach was having enough quality data from an oracle for our model to imitate. In order to account for this, we would sample the continuous space to make it discrete. Part of the action space is continuous ( and ).

We modeled our imitation approach in a similar fashion: we would build an image classifier that received the past 3-5 image frames as input and output the optimal action for the next frame. The premise of their research was to successfully apply imitation learning to play video games in the simplest possible way. However, the simplicity of the approach in Super Smash Bros was appealing. This aligned with what was taught in the lectures and seemed like an excellent solution for common problems associated with imitation learning. The MarioKart based approach leveraged DAgger to correct for deviation from the oracle. The imitation solution leveraged information from research in its application in Mario Kart and on Super Smash Bros.

Our team split into two sects: one working on a vision and controller approach and the other working on imitation AI. In this writeup, we will go over the different approaches we explored for this assignment and our development process. This would allow for one kart to play defense or pick up the puck the team kart might lose. From our experience in playing on a team in Rocket league, we also wanted to keep a certain distance between our 2 karts. Our goal was to keep the puck close to the opponent's goal as much as possible, in order to increase the amount of chances to score. Taking from our real world experience, we attempted to bring a simplistic approach to having our AI score. The hockey game we simulate is very similar to the popular game, Rocket League, which several of us have played.

Supertuxkart ice hockey simulator#

Gameplay was done in a modified racing simulator built on SuperTuxKart. Red is prediction, Green is truth Introductionįor this final class project, we attempted to build an AI that would win a 2v2 hockey game against other AI players designed by classmates. Gameplay showing ball chasing, positioning, and reversal when we pass the puck. Team Members: Dale Soney, Eric Chen, Neel Kamal, Will Cray We discuss the approach we took for this project in this write up but have ommited the code to preserve the integrity of the assignment. This was a project done for UT Austin's Deep Learning Course in Fall 2019. SuperTuxKart AI UT Fall 2019 Deep Learning Final Project