C.J. 

Autonomous drones are becoming increasingly vital in fields like search and rescue, agriculture, and law enforcement. Their success relies heavily on their ability to navigate complex environments without human intervention. Traditional training methods for these autonomous robots use batch reinforcement learning, which depends on large datasets of simulated environments that robots can navigate using trial and error. These datasets are often hard to customize and computationally taxing. Online learning offers an alternative by using continuous simulation, but it requires an efficient method to generate diverse and accurate environments in real time. This project presents a pipeline aimed at designing an efficient simulation for quadrotor training using noise-based terrain generation. The goal is to allow for scalable, efficient online learning for drones. The simulation environment is generated using Perlin noise to create smooth, randomized terrain topography. This terrain is rendered using a raymarching approach facilitated by the Python Taichi library. The simulated environment successfully rendered complex, noise-based topography, demonstrating that this component of the project was achieved as intended. Additionally, the pipeline maintained stable performance at over 600 fps indefinitely, highlighting its efficiency and effective memory management. This approach demonstrates the feasibility of a continuous simulation powered by procedural terrain generation, opening the door for implementation in training a machine learning model. Future steps will focus on adding details to the simulation to generalize to real-world environments, as well as refining the rendering system to improve clarity and efficiency.