HAZARD-RUSH: A Rhythm Racing Game and AI Agent System Based on Deep Q-Network

HAZARD-RUSH is a 2D rhythm racing game that combines rhythm game mechanics with deep reinforcement learning. It uses the Godot engine and C++ GDExtension to build an AI bridging system, demonstrating the application of the DQN algorithm in a real-time game environment. Initially developed as a case study for the COSC 304 (Introduction to Artificial Intelligence) course, this project has both game experience and AI technology demonstration value.

HAZARD-RUSH is an innovative 2D rhythm racing game that combines traditional rhythm game mechanics with deep reinforcement learning technology to create a unique competitive experience. The core concept revolves around "rhythm-driven" racing—players need to maintain speed while performing precise operations according to the music rhythm, which not only increases the game's challenge but also provides rich state spaces and decision-making scenarios for AI agent training. Project source: Original author Oumazshin, published on GitHub (link: https://github.com/Oumazshin/HAZARD-RUSH), release date 2026-05-30.

Game Engine: Godot

The Godot engine is adopted for its lightweight, open-source, and extensible features, making it suitable for independent development and academic projects. Its node system facilitates object management, and the built-in 2D rendering ensures smooth visual performance.

AI Bridging Layer: C++ GDExtension

The AI bridging system is implemented via GDExtension, acting as a translator between game logic and AI algorithms. It converts game states (player position, obstacles, rhythm timing, etc.) into numerical values understandable by neural networks, while translating AI decisions into game operations—balancing performance and flexibility.

DQN Algorithm Implementation

The core AI algorithm is Deep Q-Network (DQN), which addresses the limitations of traditional Q-learning in high-dimensional state spaces. The AI agent needs to learn abilities such as timing judgment, path planning, risk trade-offs, and competitive strategies.

Rhythm System Integration

Rhythm judgment is deeply integrated into core gameplay: acceleration, steering, and stunts need to sync with the background music rhythm. Successful synchronization grants speed bonuses, while mistakes lead to deceleration or loss of control—creating a "flow state" experience and requiring the AI to handle temporal information (predicting rhythm points).

Strategic Depth of Racing Elements

It includes classic racing elements like item collection, shortcut selection, and opponent interference, but trigger timing is tied to rhythm (e.g., the optimal use of acceleration items is at specific beat points, and shortcut entrances only open during specific music segments), providing rich reward signals for AI training.

State Space Design

The DQN agent receives multi-dimensional state information: spatial information (player/opponent position, track boundaries, obstacles), temporal information (current beat position, countdown to next rhythm point), game state (speed, item inventory, energy value), and historical information (past operation records and rhythm hit status).

Action Space Definition

Discrete action space: basic movement (accelerate, brake, turn left, turn right), rhythm actions (trigger rhythm judgment at specific times), and special operations (use items, attempt shortcuts, defensive driving)—balancing expressive power and training feasibility.

Reward Function Engineering

Composite reward structure: progress reward (distance traveled), rhythm reward (hitting rhythm points), survival reward (keeping the vehicle intact), competitive reward (overtaking opponents), and efficiency penalty (unnecessary operations/resource waste)—guiding the agent to learn efficient strategies.

Value as a Teaching Case

As a COSC304 course case, it demonstrates the transformation of AI theory into a runnable system, helping students understand Markov decision process modeling, neural network function approximation, stabilization techniques like experience replay/target networks, and the impact of hyperparameter tuning.

Reference Paradigm for Game AI

It provides a new data-driven idea for game AI: reinforcement learning agents can discover strategies humans haven't thought of, offering tools for game balance testing and difficulty adjustment. The C++ GDExtension bridging layer architecture serves as a reference for Godot developers to access high-performance external AI libraries.

Real-Time Performance Optimization

Addressing neural network inference performance bottlenecks: using lightweight network architectures, asynchronous inference mechanisms to avoid blocking the main loop, and leveraging the C++ performance advantages of GDExtension to reduce interpreted language overhead.

Training Stability

Adopting industry best practices: experience replay pools to break sample correlation, regular updates of target networks to reduce oscillations, reward scaling and clipping to prevent gradient explosions, and gradual decay of ε-greedy strategies to balance exploration and exploitation.

Multi-Agent Considerations

Currently focused on single agents, but the nature of racing implies multi-agent potential. In the future, multi-agent reinforcement learning (MARL) can be introduced to train agents that adapt to dynamic opponents.

Future Directions

• Algorithm upgrades: Double DQN (address Q-value overestimation), Dueling DQN (separate state value and action advantage), Prioritized Experience Replay (sample by importance), Noisy Nets (parameter noise instead of ε-greedy);
• Multi-modal input: Introduce visual input (game screen as CNN input) to reduce manual feature engineering;
• Meta-learning and transfer learning: Train general agents to adapt to different tracks/music/mode, quickly learn new environments via meta-learning, and reuse existing knowledge through transfer learning.

Conclusion

HAZARD-RUSH demonstrates the cross魅力 of game development and AI research—it is both a fully functional game and a living RL experiment platform. Open-source code and a clear architecture provide entry points for AI learners and developers, inspiring the game industry to think about the application potential of ML technology. In the future, more such projects will blur the line between games and AI research platforms, promoting mutual progress between entertainment and education.