Adaptive control of twin rotor system based on two-stage reinforcement learning with quantile-guided trust region

Operating-condition changes can degrade dual-channel active disturbance rejection control (ADRC) on a twin-rotor system. This study implements an artificial intelligence method, Quantile-Guided Trust Region Adaptive Control through Two-Stage Soft Actor-Critic-to-Deep Deterministic Policy Gradient Learning (Q-TRACT), for online ADRC self-tuning. Return-gated soft actor-critic (SAC) actions bound a 12 -dimensional parameter domain through componentwise 0.1 − 0.9 empirical quantiles; projected deep deterministic policy gradient (DDPG) learns within it. Union Projection enforces an undershoot-informed feasible voltage set before actuation. Archive-admission diagnostics, order-statistic sensitivity, and a conditional local model-to-parameter mismatch bound characterize transfer. A coupled-model audit verifies a strict common-Lyapunov certificate for a post hoc screened core. All five independent runs achieved two-channel settling; control-metric coefficients of variation were 7 % − 12 % . In a fixed-archive, fixed-seed comparison, untrimmed bounds increase pitch/yaw settling times by 59.2 % / 60.0 % and integral absolute error (IAE) by 48.7 % / 53.2 % relative to 0.1 − 0.9 bounds. After single-condition training, the selected policy was frozen and evaluated on a 100 -combination target grid consisting of 99 unseen joint pairs plus the training pair, and on complex references, disturbances, and Quanser Aero experiments. Relative to an architecture-matched fixed-parameter ADRC, Q-TRACT reduces the largest pitch and yaw steady-state errors by 96.8 % and 97.5 % , respectively, and the largest yaw overshoot by 87.1 % , although its largest pitch overshoot is 44.7 % higher. Ramp experiments show 49.5 % − 77.6 % lower IAE than Manual Tuning. Crossed ablations show no universal metric dominance and a controller-dependent effect of Union Projection. Q-TRACT provides ADRC self-tuning with execution-layer input admissibility across the tested conditions.