
Several recent works have demonstrated that Nonlinear Model Predictive Control (NMPC) schemes can be effectively designed using the Linear Parameter Varying (LPV) framework. In particular, LPV MPC have been shown as a powerful alternative to standard NMPC, reducing the overall involved computational stress while ensuring (near) optimal performance. Despite recent advances, many practical applications consist of algorithm structures to select the optimisation problem to be solved at each sampling instant (w.r.t. the scheduling variable), and still rely on proprietary solvers to solve these involved Quadratic Programs. In this paper, we propose the first dedicated open-source solver for NMPC using the LPV approach: LPV-PILOT, which is available online for testing. In particular, our software comprises an automatic solution to implement LPV MPC, preparing and solving the QP to be solved at each sampling period. The solver is based on consolidated convex optimisation techniques, i.e. interior-point, coupled to quasi-LPV embeddings,constraints tightening, warm-start and fallback strategies. The rationale on using conventional methods is for the solver to be widely replicable in industrial settings, being based on a few simple algorithmic structures, and thereby making certification easier. Furthermore, the developed scheme is open-source, compatible with MATLAB and its simple algorithmic structure facilitates embedded C/C ++ implementations. To demonstrate the effectiveness of LPV-PILOT in engineering, several validation results are presented: a numerical simulation example of a twin-rotor system is used to discuss the numerical aspects of the solver; a hardware-in-the-loop helicopter landing problem is presented, for which the solver is embedded in an ESP32 microcontroller; and a thorough comparison of LPV-PILOT with a widely used NMPC toolbox from the literature is included: CasADi running ipopt. Our results indicate the numerical competitiveness of LPV-PILOT, with highlights to its simple structure and standard convex techniques. Without involving any multiple-shooting discretisation or real-time iterations, our scheme is shown to generate NMPC laws in the microsecond range in low-cost hardware environments.
Accurate real-time parameter estimation is important for achieving higher performance control of permanent-magnet synchronous motors (PMSMs). Existing estimation methods often involve parameter-specific designs and heuristic observer tuning, which limits practical implementation. This paper proposes a unified observer structure with parameter-specific excitation conditions for real-time estimation of stator resistance, inductance, and permanent-magnet flux linkage. A tuning methodology is developed to provide a systematic observer gain selection procedure through pole assignment and dynamic response analysis. The proposed framework enables a consistent design procedure across different electrical parameter estimators while maintaining implementation simplicity. Experimental results under various operating conditions demonstrate the effectiveness and practical applicability of the proposed method.
This paper evaluates a two-time-scale architecture for remote reinforcement learning (RL) tuning of a proportional–integral–derivative (PID) controller without placing network communication in the embedded sensor-to-actuator loop. An Arduino Uno R4 Wi-Fi executes the local PID on a ball-in-tube apparatus, while a remote twin delayed deep deterministic policy gradient (TD3) supervisor proposes bounded gains over Wi-Fi. The embedded controller remains operational during communication delay and reverts to calibrated fallback gains after disconnection. A frozen TD3 policy was evaluated from deliberately suboptimal gains in ten matched physical blocks under randomized supervisory delays of 0, 30, 60, and 100 ms, with nonzero jitter. Each block compared calibrated fixed PID, timing-aware RL, and the same actor with timing observation and delay-dependent damping disabled. All 30 runs completed without a safety-zone event or abort, and all 870 commands were acknowledged. A 6,000-sample identification run and discrete-time Jury analysis established a local model-conditional gain region containing every confirmatory command. The combined timing-aware mechanism robustly reduced normalized gain-command total variation by 19.5 p<0.001 ). Tuning-window mean absolute tracking error was 5.4 p=0.031 ), a suggestive result that did not meet a conservative Bonferroni threshold for the two primary comparisons. After convergence, timing-aware RL and calibrated fixed PID had descriptively similar errors ( p=0.857 ), although this was not an equivalence test and RL produced more pulse-width modulation saturation. The results support remote supervisory PID tuning feasibility on the studied platform while characterizing timing-aware damping as a conservative update regularizer, not compensation for delay in the local control loop.
This paper presents a continuous-time switching control strategy tailored to mitigate chattering in switched affine systems under digital implementation. Inspired by discrete-time practical stability concepts, the proposed method introduces an optimized geometric relaxation framework that restructures the state-space geometry around the equilibrium point. Specifically, the control law optimizes the volume of an invariant ellipsoidal attractor while systematically relaxing the stability constraints along non-critical directions, thereby allowing bounded, controlled limit cycles to manifest without triggering high-frequency switching. The switching rule and the robust invariant set boundaries are simultaneously synthesized via a set of linear matrix inequalities (LMIs). The methodology is validated through two numerical benchmarks: a DC machine speed control and a DC–DC buck converter under parametric uncertainties. Simulation results at 100 kHz demonstrate a significant reduction in the effective switching frequency—achieving 143.6 Hz for the DC machine and 1.19 kHz for the buck converter—while strictly maintaining bounded tracking errors. These findings demonstrate that the proposed technique can successfully reduce the chattering condition in switched systems, increasing performance and broadening the practical applicability of switched control systems.
To address the problem of radar-evading path planning for sea-surface unmanned aerial vehicles (UAVs) under constant-altitude and constant-speed constraints, conventional methods simplify the UAV as a point mass, ignore the dynamic relationship between radar cross section (RCS) and flight attitude, and replace the realistic radar detection model with geometric no-fly zones. Consequently, these approaches suffer from physical distortion and poor engineering practicality. To overcome these limitations, this paper proposes a path planning algorithm that integrates attitude-modulated RCS with proximal policy optimization (AMRCS-PPO). Under strict altitude hold enforced by a height autopilot, the UAV actively modulates its RCS using pitch attitude while controlling the horizontal path via yaw attitude. A dynamic RCS model coupled with yaw and pitch attitudes and a continuously differentiable detection probability model are constructed based on the radar equation. A reward function centered on minimizing cumulative detection probability is designed, and the PPO algorithm is adopted for joint attitude and path decision-making. Simulation experiments compare the proposed algorithm with A*, RRT*, and conventional PPO. Results demonstrate that the proposed algorithm significantly outperforms A* and RRT* in terms of both average and maximum detection probability. Compared with conventional PPO, the average detection probability is reduced by 79.5