// the find
drkostas/Minecraft-AI
A Reinforcement Learning agent that learns how to solve maze missions in Minecraft.
A university project that trains a PPO agent to navigate Minecraft mazes using Microsoft's Malmo library. It takes raw pixel input through a CNN and tests a handful of action space and dimensionality reduction variations. Aimed at students or researchers looking for a working Malmo + stable-baselines3 starting point, not a production RL framework.
The dynamic fire mechanic adds genuine environment complexity beyond static mazes — the agent has to learn avoidance, not just pathfinding. The action space ablation (6-action vs 4-action) is actually documented with results, which is more rigorous than most class projects. 160 maze XML files gives a decent curriculum for generalization testing. The setup script is explicit enough that someone on Ubuntu 22.04 can reproduce it without guessing.
Python 3.6 and CUDA 11.0 are both end-of-life; Malmo itself is largely unmaintained and getting it running on anything other than the exact tested environment is a project in itself. No saved model weights are included, so you have to burn 5 hours of GPU time just to reproduce the paper's results. PCA failure is noted but not diagnosed — the conclusion is 'PCA bad, tune it maybe', which is not useful. The repo is a snapshot of a class assignment; there's been no activity since 2023 and the 8 forks suggest minimal adoption.