· Game Dev · 1 min read
EmoGalaxy 2: Teaching NPCs to Feel with Unity ML-Agents
A deep-dive into my thesis project — training autonomous game agents to exhibit emergent emotional behaviour using Unity ML-Agents and reinforcement learning.
Overview
EmoGalaxy 2 is my university thesis project exploring how reinforcement learning can produce emergent, emotionally-rich NPC behaviour inside a real-time game environment.
The Problem
Traditional game AI uses finite state machines or behaviour trees that are hand-authored — emotions are scripted, not felt. The goal was to see whether an agent trained purely on reward signals could develop behavioural patterns that look emotional to human observers.
Architecture
- Unity ML-Agents Toolkit — training environment and policy inference
- PPO (Proximal Policy Optimization) — the core RL algorithm
- Observation space — positional data, proximity to other agents, recent health delta, sound cues
- Reward shaping — sparse rewards for survival and cooperation; penalty for isolation
Key Findings
Training for ~50M steps produced agents that exhibited clustering under threat (fear-analogue), competitive aggression near resources (anger-analogue), and exploratory behaviour during calm periods (curiosity-analogue). None of these were explicitly programmed.
What I’d Do Differently
Curriculum learning from the start would have cut training time by roughly 40%. The initial reward landscape was too sparse, causing the agents to plateau early.
Full thesis write-up and training logs available on request.
