Comparative Evaluation of Classical, Predictive, and Learned Control Paradigms for Mars Rotorcraft Under Authority and Uncertainty Constraints

This paper presents a rigorous, like-for-like comparison of three low-level flight control paradigms for Mars rotorcraft: classical cascaded PID control, Model Predictive Control (MPC), and deep reinforcement learning (RL-CEM). Implemented on a 1.8 kg coaxial airframe model in MuJoCo across 10,206 simulated episodes, we evaluate tracking performance, robustness, and computation constraints across a sweep of thrust-to-weight margins ($\mu$). Our findings challenge prevailing assumptions: under sampled uncertainty, an authority-adapted classical loop achieves lower median tracking error than MPC while requiring significantly less computation. Furthermore, solver iteration counts can be cut five-fold without significant performance loss, whereas prediction horizon truncation severely degrades tracking. Finally, domain-randomized RL policies exhibit severe out-of-distribution generalisation failure when transferred across gravity regimes. Contents & Materials Included Manuscript (.pdf): Complete research study including methodology, uncertainty models, experimental results, and sensitivity analyses. Simulation Data & Artifacts: Built upon MuJoCo simulation frameworks evaluating classical PID, predictive, and learning-based control laws.