Rocket Lander RL Visual Lab

Scaffold created: 2025-12-15 21:11:25

A modular reinforcement-learning + visualization project where a DQN agent learns to land a rocket on a pad. Focus: watchable learning (Q-values, intent vectors, replays, curriculum), no sound.

Structure

• rocket_lander/ core package
• main.py entrypoint (to be implemented)
• outputs/ checkpoints + run exports
• tests/ future test suite

Headless training + visualization

Run training without Pygame rendering via headless_train.py, then optionally spin up a short visualization of the trained policy:

# Train for the configured number of episodes (headless)
python headless_train.py --episodes 500

# Skip training and visualize the latest checkpoint for 3 episodes
python headless_train.py --skip-train --visualize --visualize-episodes 3

Next steps

1. Move your current single-file lander into:
   • rocket_lander/terrain.py
   • rocket_lander/lander.py
   • rocket_lander/agent.py
   • rocket_lander/trainer.py
   • rocket_lander/game_loop.py
2. Keep constants in rocket_lander/config.py.
3. Add features incrementally: particles, Q-bars, showcase, replay, ghost.