Playing Flappy Bird Based on Motion Recognition Using a Transformer Model and LIDAR Sensor

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• The Flappy Bird game is a popular benchmark for reinforcement learning control tasks.
• Traditional approaches often rely on visual input (game rendered image).
• There is a need for alternative sensory inputs and methods for robust agent navigation.

Purpose of the Study:

• To develop a reinforcement learning agent capable of navigating the Flappy Bird game.
• To investigate the efficacy of using LIDAR (Light Detection and Ranging) sensory input, processed via ray casting, for agent control.
• To enhance obstacle avoidance by analyzing temporal context and proximity to obstacles.

Main Methods:

• Utilized a transformer neural network architecture for predicting Q-values.
• Employed reinforcement learning techniques to train the agent.
• Implemented a ray casting method to represent LIDAR sensory input, focusing on temporal measurements.
• Optimized agent behavior by considering the degree of approach to obstacles.

Main Results:

• The agent successfully learned to navigate the Flappy Bird environment.
• The agent effectively utilized ray casting measurements to avoid collisions with obstacles.
• The proposed model demonstrated substantial performance improvements over related approaches.

Conclusions:

• Transformer neural networks combined with LIDAR ray casting provide an effective alternative to visual input for reinforcement learning agents.
• This approach enhances obstacle avoidance capabilities by leveraging temporal sensory data.
• The method shows promise for application in real-world navigation and control scenarios.