Skip to content

README.md

Latest commit

 

History

History
65 lines (55 loc) · 3.15 KB

README.md

File metadata and controls

65 lines (55 loc) · 3.15 KB
Raw

Federated-Imitation-Racing

Description

Federated-Imitation-Racing is a project that explores supervised federated learning applied to autonomous driving in the context of Mario Kart Wii.
This project serves as a proof-of-concept for federated learning. Player actions—such as steering, acceleration, and item usage—are recorded as labeled datasets.
While the project currently does not use multiple physical clients, it simulates multiple clients by training separate models on local datasets.
Federated learning techniques then allow these models to collaboratively improve by sharing model weights or gradients, rather than raw gameplay data, preserving privacy while improving performance.

Features

  • Federated Learning: Train models across multiple clients without centralizing sensitive gameplay data.
  • Imitation Learning: Learn to drive by mimicking players.
  • Privacy-Preserving: Only model are shared; raw data remain local.
  • Simulation-Compatible: Works with Mario Kart Wii gameplay data recorded from Dolphin emulator or other sources.

Installation

Prerequisites

Before running the project, make sure you have the following installed on your system:

  1. Docker Desktop

  2. Dolphin Emulator (with Python scripting support)

    • Required for recording gameplay and running automated scripts.
    • Clone: https://github.com/Felk/dolphin
    • This is a branch with Python Scripting support.
    • Place the Dolphin Installation in dolphin/.
    • Place any custom Dolphin scripts under client/dolphin/Scripts/.

Training models

It is assumed the same map, vehicle and character will be played anytime data is gathered.
  1. Record your game data
    • Start Dolphin.
    • Load Mario Kart Wii (PAL).
    • Run "record_training_data.py" through the scripting interface of Dolphin.
    • Play some races.
  2. Run the compose 
    • docker compose build
docker compose up
    • This will train and send a model to the central server.
  3. Repeat at least 3 times
    • After 3 models are sent, the server will aggregate them together.
    • At the same time, the centralised model will be trained on all that data together at once.
    • Then it will reset its data too to make a fair comparison.
  4. Gather validation data
    • Run "record_validation_data.py" through the scripting interface of Dolphin.
    • Play some races.
  5. Run comparison
    • Run "compare.py".
    • This pipes the validation data through each model and compares performance.

Playing Model Live

  1. Setup game
    • Start Dolphin.
    • Load Mario Kart Wii (PAL).
    • Get ready to play race.
  2. Run model live
    • Run the "run_model_live_server.py" script externally.
    • Run the "run_model_live_emu.py" through the Dolphin scripting interface.
    • Set the model based on MODEL_NAME in the server script.
    • A save state will be loaded and you can watch the named model play live.