-
Notifications
You must be signed in to change notification settings - Fork 25
Home
This documentation might see other changes in the near futur, anyways here is Saltie's Wiki For Ya ;).
Our purpose is to make an open-source framework to enable the creation of Rocket League bots using machine learning. Feel free to download the repository and test it yourself, but never republish it without permission of the owner! This framework uses the features of RLBot to get its inputs and give its outputs. As there is no consensus as to the best way to train a machine learning bot for Rocket League, multiple avenues are currently being explored.
On this wiki you can find info about the framework, instructions for how to set up your own bot, some information about the inputs from the bot, and later on probably a lot more!
This bot mainly learns through mimicry. It learns to play Rocket League by learning to output the same controls as other bots. Thus, we record games where other bots play against each other to obtain sufficient training data. Recording games also makes the data easier to manipulate (such as if new features were made based on existing ones). Instead of having one game running in realtime, Saltie trains on multiple games simultaneously, hugely improving its learning capacities.
If you want to train in a reinforcement manner, Saltie's framework provides a good start for integration with RLbot. Supervised learning is the current priority because it gives us a baseline, from which the bot can start improving itself.
There are multiples bot config files stored in root(saltie.cfg, saltie2.cfg, saltieRNN.cfg, saltiek.cfg, tutor_bot.cfg,... ) because depending on which model you want to use there are different sets of parameters you would need to input.(layer numbers, layer size, is_learning, ...).
All the code used by Saltie are stored in the bot_code folder.
The agents folder contains bots you would train copying/record play-troughs.
Everything else is related to RLBot, give it a look.