From Pygame to Trained Agents with Open-Source Tools

Promo Slides for Workshop Tutorial for OpenSouthCode 2026 Submission

Transcript

1. BuildYourOwn RLPlayground From Pygame to Trained Agents with Open-Source Tools Málaga · OpenSouthCode 2026 · 2h Workshop ▷ @AlgorithmicMatters
2. The Problem Everyone runs someone else’s RL environments. Almost nobody learns to build their own. pip install gym Run CartPole demo Copy a training script ??? What next ??? Design your own game world Wrap it as a Gymnasium env Train agents & watch them learn This workshop ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
3. The Open-Source Stack Everything we use is free, open-source, and community-maintained. Pygame Game & simulation Gymnasium Standard RL API Stable- Baselines3 RL algorithms TensorBoard Visualization NumPy PettingZoo multi-agent Godot (next step) You leave with a forkable MIT-licensed repo ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
4. What We’ll Build The Game (worked example) • Agents move on a 2D map • Signal sources that must be covered • Agents have limited vision — no god’s-eye view • Cooperation is required to cover all signals The theme is a starting point. Reskin it as: → Drones mapping a disaster zone → Robots in a warehouse → Characters in a cooperative roguelike → Your idea here! SwarmSignal — Pygame • Agent • Signal ▪ Obstacle ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
5. Workshop Roadmap — 2 Hours 0:00 0:30 1:00 1:30 2:00 30 min Simulation Agents, obstacles, collisions, render loop + human play mode 30 min Gymnasium reset(), step(), obs/action spaces, seeding, reproducibility 30 min SB3 Training PPO, reward shaping, watch agents learn, TensorBoard 15 min Next Steps Partial observability, communication, Godot bridge, OSS repo Bring your laptop · Python required · No RL experience needed ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
6. A Taste of the Code The Gymnasium wrapper import gymnasium as gym from gymnasium import spaces import numpy as np class SwarmEnv(gym.Env): metadata = {" render_modes ": ["human", "rgb_array"]} def __init__(self , n_agents =4, grid_size =20): super ().__init__ () self. action_space = spaces.Discrete (5) self. observation_space = spaces.Box( low=0, high=1, shape =( n_agents , 8), dtype=np.float32 ) def reset(self , seed=None , **kw): super ().reset(seed=seed) # ... init world state return obs , info def step(self , action): # ... physics + rewards return obs , reward , done , truncated , info Training with Stable-Baselines3 from stable_baselines3 import PPO from stable_baselines3 .common.env_util \ import make_vec_env # One line to vectorize env = make_vec_env ( "SwarmSignal -v0", n_envs =4 ) # Train model = PPO( "MlpPolicy", env , verbose =1, tensorboard_log ="./ tb_logs/" ) model.learn( total_timesteps =100 _000) # Watch the result ! obs = env.reset () for _ in range (1000): action , _ = model.predict(obs) obs , reward , done , info = \ env.step(action) env.render () ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
7. The Payoff: AI agent learns in your environment Reward Training steps Agents learn to cooperate! You built the game. You trained the AI. You own the whole stack! ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
8. What You Take Home Template repo (MIT license) Fork, modify, publish — it’s yours Gymnasium API mastery The standard interface for all RL SB3 training pipeline PPO on your own environment Clear extension path Multi-agent, Godot, your own theme ✓ Python basics is the only prerequisite ✓ No RL experience needed ✓ Bring a laptop ▷ @AlgorithmicMatters OpenSouthCode 2026 · Málaga
9. Join the Workshop Build · Train · Open-Source Where: OpenSouthCode 2026 — Málaga Format: 2-hour hands-on workshop Web: opensouthcode.org/conferences/opensouthcode2026 YouTube: @AlgorithmicMatters — subscribe for updates! Code: Repo link coming soon Built with open-source. Shared as open-source. That’s the whole point.