Stick Fighter — AI Training Platform

A 2D stick figure fighting game built with Pygame. Play against scripted AI, train a clone that learns your style, then face an adaptive AI that exploits your patterns.

Quick Start

pip install -r requirements.txt
python main.py

Controls

Key	Action
A / Left Arrow	Move left
D / Right Arrow	Move right
W / Up Arrow	Jump
J	Punch (fast, low damage)
K	Kick (slower, more damage)
L	Uppercut (slow, high damage)
Space	Block
ESC	Quit
R	Restart (after game over)

Game Modes

# VS scripted AI (default) — generates training data
python main.py
python main.py --p2 scripted --p2-aggression 0.8

# VS random AI
python main.py --p2 random

# VS your trained clone (imitation model)
python main_vs_clone.py
python main_vs_clone.py --temperature 0.5
python main_vs_clone.py --deterministic

# VS adaptive AI (blends imitation + counter model)
python main_vs_adaptive.py
python main_vs_adaptive.py --temperature 0.8
python main_vs_adaptive.py --start-alpha 1.0

Training Pipeline

1. Record gameplay

Play sessions against the scripted AI. Recordings are saved automatically to recordings/.

python main.py

2. Train the imitation model

Trains an LSTM to predict your actions from gameplay recordings.

python -m training.train_imitation

3. Train the PPO counter model

Trains a PPO agent to exploit your predicted patterns by playing against the imitation model.

python -m training.train_ppo

4. Play against the adaptive AI

Blends both models with a difficulty schedule that ramps up each round.

python main_vs_adaptive.py

File Structure

File	Purpose
config.py	All tunable parameters (damage, speed, screen size, etc.)
actions.py	Action enum — the shared language between human and AI
fighter.py	Fighter state machine (movement, attacks, physics)
controllers.py	Human keyboard input + scripted AI controllers
game_utils.py	Collision detection and game state building
game_loop.py	Shared game loop used by all game modes
renderer.py	Stick figure drawing and UI
recorder.py	Logs every frame to .jsonl for training
env.py	Gymnasium-style wrapper for RL training
main.py	VS scripted/random AI
main_vs_clone.py	VS your trained imitation clone
main_vs_adaptive.py	VS adaptive difficulty AI
training/imitation_model.py	LSTM model for imitation learning
training/imitation_controller.py	Game controller for imitation model
training/train_imitation.py	Imitation training script
training/dataset.py	Loads recordings into training sequences
training/ppo_model.py	Actor-Critic network for PPO
training/train_ppo.py	PPO counter model training script
training/ppo_controller.py	Game controller for PPO counter model
training/difficulty_blender.py	Blends imitation + counter with alpha schedule

Recording Format

Each play session creates a .jsonl file in recordings/. Each line is one frame:

{
    "frame": 142,
    "p1_state": [0.25, 1.0, 0.85, ...],
    "p2_state": [0.75, 1.0, 0.70, ...],
    "distance": 0.50,
    "p1_action": 4,
    "p2_action": 1,
    "p1_hp_delta": 0,
    "p2_hp_delta": -8,
    "round": 1,
    "round_timer_frac": 0.35,
    "p1_wins": 1,
    "p2_wins": 0
}

Using the Gym Environment

from env import FightEnv

env = FightEnv(opponent_type="scripted", render=False)
obs = env.reset()

for step in range(10000):
    action = 4  # punch
    obs, reward, done, info = env.step(action)
    if done:
        obs = env.reset()