Reinforcement learning with formal performance metrics for quadcopter attitude control under non-nominal contexts

This project offers a complete framework for designing and testing reinforcement‑learning (RL) attitude controllers for quadcopters that must fly in non‑ideal conditions. Using the Crazyflie 2.0 micro‑quadcopter’s physical model, we simulate tough situations such as partial loss of rotor power and strong wind gusts. To evaluate the controllers more thoroughly than just using cumulative reward, we add a solid Signal Temporal Logic (STL) formulation to the testing process. STL lets us measure transient response, steady‑state error, and safety limits in a quantitative way. The study explains how we chose deep‑neural‑network architectures, defined the state‑observation space, and tuned hyper‑parameters specifically for stabilizing multi‑rotors. Through many simulations, we compare the robustness of

the trained RL policies with classic control methods when faced with severe environmental and structural faults. The implementation gives clear guidance for building highly resilient flight‑control systems and provides a rigorous method for validating neural‑network‑based controllers in safety‑critical aerospace applications. This work connects practical machine‑learning results with formal control theory.

Message: Hi FE, I need help with “Reinforcement learning with formal performance metrics for quadcopter attitude control under non-nominal contexts” in “Aerospace Engineering”