keyboard_arrow_up
A Dynamic Simulation Platform to Train and Control Rocket Landings using Unity ML-Agents and Reinforcement Learning

Authors

Haoran Jin1 and Andrew Park2, 1USA, 2California State Polytechnic University, USA

Abstract

This project addresses the challenge of simulating rocket landings across different planetary environments by using Unity ML-Agents to train AI models [1]. The reusability of rockets, critical for space exploration, requires precise control and adaptability to varying gravitational conditions. We proposed a solution combining AI-driven controls with interactive user input to create a flexible and realistic rocket landing simulator. The methodology employed machine learning to develop models capable of handling complex control tasks, using reinforcement learning to adapt to the distinct environments of Earth, Mars, and the Moon. Experiments centered on evaluating the model's ability to adjust and perform within each environment, analyzing how critical rocket parameters, such as mass and thrust, influenced performance across varied gravitational and atmospheric conditions. This approach provided insights into the model’s adaptability and optimization potential for diverse extraterrestrial applications. [2]. The most significant finding was that the AI performed well on Earth and the Moon but required further tuning on Mars due to faster descent speeds [3]. Our approach provides an engaging and educational platform for studying reusable rocket technology, making it a valuable tool for both academic and practical applications.

Keywords

Machine Learning, Rockets, Landing, Reinforcement Learning

Full Text  Volume 14, Number 21