Abstract:
Quadcopters have already proven their effectiveness in both civilian and military applications. Their control, however, is a difficult task due to their under-actuated, highly
nonlinear, and coupled dynamics. Most quadcopter autopilot systems utilize cascaded
control schemes, where the outer loop handles mission-level objectives in 3D Euclidean
space, and the inner loop is responsible for stability and control. Such complex systems
are generally operated using PID controllers, which have demonstrated exceptional performance in multiple scenarios, such as obstacle avoidance, trajectory tracking and path
planning. However, tuning their gains for nonlinear systems using heuristics or rulebased methods is a tedious, time-consuming and difficult task. Rapid advances in the
field of computational engineering, on the other hand, have paved way for intelligent
flight control systems, which have become an important area of study addressing the
limits of PID control, most recently through the application of reinforcement learning
(RL). In this dissertation, an optimal gain auto-tuning strategy is implemented for altitude, attitude, and position controllers of a 6 DoF nonlinear drone system using a deep
actor-critic RL algorithm having continuous observation and action spaces. The state
equations are derived using Lagrange’s (energy-based) method, while the drone’s aerodynamic coefficients are estimated numerically using blade element momentum theory.
Furthermore, the cascaded closed loop system’s asymptotic stability is studied using the
theory of Lyapunov. Finally, the proposed strategy is validated by simulation results,
where the gains learned by RL agents allow the quadcopter to track a given trajectory
accurately. Moreover, these optimal gains satisfy the conditions obtained through Lyapunov’s stability analysis, indicating that the RL algorithm is an extremely powerful
tool which can assess uncertainties existing within any complex nonlinear system.
