Phase-Aware Multi-Scale Encoding for Risk-Aware Twin-Actor Quadcopter Control

We introduce PhASE-RTAQC, a phase-aware multi-scale encoding framework for risk-aware twin-actor quadcopter control with Proximal Policy Optimization (PPO). A Gated Recurrent Unit (GRU) summarizes recent state history to infer a bounded continuous control-regime index, ϕt∈(0,1). This index is not defined by prescribed physical thresholds and should not be interpreted as a calibrated probability of instability. Instead, it represents a learned continuous coordinate capturing control-relevant temporal regimes from recent state history. Temporal information is further encoded using dilated 1D convolutions at multiple scales. Control actions are generated by twin Gaussian actors corresponding to nominal and recovery-oriented behaviors and are smoothly combined through a regime-dependent routing mechanism. The control-regime estimator is trained jointly through the reinforcement-learning objective, without explicit regime labels or auxiliary supervision. A distributional critic provides risk-sensitive value estimation, while PPO uses lower-tail Conditional Value at Risk (CVaR)-style Generalized Advantage Estimation (GAE) to place greater emphasis on low-return trajectories during training. Under the final simulation and actuator configuration, PhASE-RTAQC achieves, over the final 100 training episodes, a mean reward of 2345.16, a mean terminal altitude of 75.37 m, and a mean terminal position error of 0.39 m for the 75 m target. A matched comparison with an adapted Game-Theoretic Constrained Policy Optimization (GCPO) controller is performed under nominal, mild, and strong perturbations using the same simulation dynamics and evaluation protocol. The two controllers show scenario-dependent trade-offs under nominal and mild perturbations, while both degrade under strong perturbations, with GCPO maintaining stronger constraint preservation. Sensitivity tests show higher sensitivity to mass and thrust-model mismatch than to the tested observation noise and delay. These results characterize PhASE-RTAQC as a simulation-based, risk-aware control framework whose conclusions are limited to the evaluated conditions. The controller has not yet been validated through hardware-in-the-loop or physical-flight experiments, and the study does not establish formal guarantees of constraint satisfaction, closed-loop stability, policy convergence, robustness, or safe operation on a physical Unmanned Aerial Vehicle (UAV).