The increasing adoption of encapsulated motor drives in industrial automation presents a fundamental control challenge: low-level Field-Oriented Control (FOC) loops are embedded in proprietary firmware, exposing only standardized high-level command interfaces while concealing internal states, parameters, and dynamics. Model-based approaches such as Proportional–Integral–Derivative (PID) and Model Predictive Control (MPC) require explicit system identification that degrades under thermal drift, load variations, and aging. This paper presents a deployment-oriented offline reinforcement learning (RL) framework that trains compact neural network policies exclusively from interface-level feedback, without access to internal drive states or explicit parameter identification. A three-stage progressive training methodology is developed: (I) simulation pretraining via behavior cloning from an MPC expert with conservative value regularization to inject safe control priors; (II) mixed-data fine-tuning that leverages suboptimal PID trajectories as negative constraints through an anchor-margin critic loss, explicitly encoding undesirable behaviors without online interaction; and (III) real-world offline adaptation using an asymmetric-layer transfer strategy that resets input projection layers while preserving logic layers, bridging the sim-to-real perception gap. The framework is validated on two distinct embedded platforms: a bare-metal STM32F407 controlling four permanent magnet synchronous motors (PMSM) via RS485 (50 Hz control cycle), and a FreeRTOS-based APM32A407 deployed on a full-scale swerve-drive vehicle with eight brushless direct current motor (BLDC)—four drive and four pivot—communicating over CAN bus (100 Hz RL inference with 10 Hz sensor telemetry). On the large-scale platform, the RL-augmented controller achieves 40% faster convergence (6.0 s vs. 10.0 s) and 55% lower overshoot (8.3 RPM vs. 18.5 RPM) compared to a well-tuned MPC baseline, while reducing average power consumption to 6.94 W (5.0% below MPC). Across both platforms, the approach demonstrates cross-platform methodological transfer without per-platform model redesign.
This paper presents an adaptive control framework for quadcopter trajectory tracking in which all external state feedback is derived from a monocular video stream. Inspired by first-person view (FPV) piloting, the proposed system uses monocular video streamed to a remote ground station to estimate pose, identify dynamic parameters online, and compute control commands via model predictive control (MPC), while low-level attitude stabilisation is performed by the onboard flight controller in inertial measurement unit (IMU)-based angle mode. The architecture integrates ORB-SLAM3 for real-time pose estimation, an augmented-state Unscented Kalman Filter (UKF) for online estimation of internal quadcopter model parameters, and an MPC controller. All estimation and trajectory-level control computation is performed offboard, requiring only a hobby-grade quadcopter equipped with a monocular FPV camera, video transmitter, and radio-control receiver. Real-world experimental results demonstrate adaptation across quadcopter configurations and accurate tracking of multiple reference trajectories.
Path tracking control methods for mobile robots show poor applicability and limited performance in coal mine underground with complex scene interference, such as slip and unstructured pavement impacts. Therefore, a kinematic model considering slip is used as the basis, combined with the lateral response in the dynamics model, to build dynamics-kinematics hybrid model with the rotational acceleration as the virtual control input, which utilizes generalized disturbances to represent the uncertainty induced by robot slippage. Then, the extended state observer (ESO) is proposed to estimate the traveling state and generalized disturbance, which addresses the difficulty of directly measuring velocity and disturbance information. Furthermore, the hierarchical control strategy is proposed based on path tracking controller and track speed controller. The path tracking controller is developed for realizing the mobile robot kinematic path tracking, and the track speed controller is proposed for the dynamics track velocity control, which possesses excellent interference immunity and traveling control accuracy. Finally, the experimental study on the path tracking control of mobile robot is carried out in the simulated coal mine underground. The results show that the proposed method can well realize the traveling control under the complex scene interference. The RMS values of trajectory tracking errors are 5.86 cm, 8.22 cm, and 0.0972 rad, and the chattering phenomenon is obviously improved.
Electromagnetic levitation system is a core subsystem of high-speed maglev. The levitation unit forms a coupled 2-DOF rigid system, which bring strong coupling effects. Besides, external disturbances such as track irregularities and aerodynamic drag severely affect levitation stability. To address these issues, the nonsingular terminal sliding mode control (NTSMC) combined with an improved nonlinear extended state observer (NESO) is designed subject to coupling effects and external disturbances. The matrix diagonalization-based feedback decoupling method is utilized to suppress system coupling effects. A smooth saturation function replaces the traditional piecewise function to construct an improved NESO for lumped disturbance estimation, while the NTSMC is designed to achieve rapid tracking. The combined scheme avoids singularity issues and achieves finite-time convergence, as verified theoretically. Simulation results indicate that the designed controller enables satisfactory levitation performance under diverse track irregularities and aerodynamic resistance, even under high-speed conditions. Finally, the effectiveness of the proposed controller is further validated on a real high-speed maglev levitation test platform.
Organic Rankine cycle systems exhibit significant nonlinearity and are commonly affected by heat-source fluctuations and load variations, which cause frequent changes in operating conditions over a wide range. As a result, fixed prediction models have difficulty maintaining consistent prediction accuracy over the entire operating range, which lead to control performance degradation. To address this problem, this paper proposes an input-mapping model predictive control method with event-triggered adaptive mechanism(ETA-IMMPC). First, an input-mapping method is employed to compensate for unknown modeling biases using online historical input-output data. Furthermore, to address severe model mismatch caused by wide-range operating-condition variations, an event-triggered adaptive model updating mechanism is constructed. When the degree of model mismatch exceeds the compensation capacity of the input-mapping method, online model updating is triggered. The model is then updated using historical data, and input-mapping model predictive control is further implemented based on the updated model, thereby achieving a balance between control performance and computational burden. Finally, the proposed method is verified on an Aspen Plus/Simulink co-simulation platform under heat-source disturbances and load variations.
This paper investigates the trajectory tracking control of free-floating space manipulators (FFSMs) subject to a time-varying input delay using the fully actuated system (FAS) approach. By leveraging momentum conservation, the FFSM dynamics is formulated into an FAS model. This enables the FAS approach to cancel out open-loop nonlinearities, yielding a linear time-invariant closed-loop system with a parametrically assignable eigenstructure. To compensate for the input delay, an auxiliary predictor is designed. Furthermore, a delay estimation scheme and an improved predictor are developed to address the practical implementation issues caused by teleoperation-network effects, including bounded random jitter, packet loss, and out-of-order data arrivals. Numerical simulations are conducted on a two-degree-of-freedom planar FFSM performing a servicing-task trajectory, where the command channel is modeled as a relay-assisted ground-to-space teleoperation link with time-varying propagation delay and packet-level perturbations. The results demonstrate that the proposed framework maintains stable tracking performance under this networked teleoperation scenario.
Building a reliable model is the cornerstone of ensuring the safe operation of complex systems engineering. The belief rule base (BRB) is extensively employed in this field because of its ability to effectively integrate prior knowledge. However, current BRB modeling focuses mainly on interpretability, while research on knowledge reliability remains insufficient. Therefore, this paper proposes a BRB with multi-stage prior knowledge reliability integration (BRB-PKR) for reliable modeling of complex systems. Firstly, in response to the potential cognitive uncertainty of single source prior knowledge, a multi-source prior knowledge fusion mechanism is introduced to initialise prior knowledge reliability and build the initial BRB model with prior knowledge reliability. Secondly, the belief distribution in the reasoning engine is reallocated based on rules reliability to ensure that the reliable rules play a leading role in evidence fusion. Additionally, an optimization method with reliability constraints is built to maintain the model reliability during parameter optimization. Lastly, two case studies are used to confirm the effectiveness of the method. The experimental outcomes indicate that this method enhances prediction accuracy while preserving model reliability, quantifies the reliability of output results, and offers a transparent and reliable solution for reliable modeling of complex systems engineering.
In uneven terrain, complex and undulating terrain structures can weaken a robot’s ability to assess traversable areas, reduce exploration efficiency, and increase motion risks during path execution. Therefore, autonomous exploration by ground robots in uneven environments remains highly challenging. This paper proposes an autonomous exploration framework that integrates a terrain-aware node graph with deep reinforcement learning. First, a terrain-aware non-uniform node graph is constructed based on a traversability map, organizing local terrain complexity, regional reachability, and spatial connectivity into a graph-structured representation for high-level exploration decision-making, thereby enhancing the ability of the environmental representation to capture uneven-terrain characteristics. On this basis, deep reinforcement learning is used to learn an exploration target selection policy, and terrain information is incorporated into node utility evaluation, value estimation, and policy optimization, enabling the robot to simultaneously consider unknown-area rewards, terrain traversability conditions, and long-term exploration returns during exploration. Simulation results show that the proposed method completes exploration tasks in uneven environments with shorter exploration paths and less exploration time than the comparison methods. Real-world experiments further verify the feasibility of the framework in practical uneven-terrain scenarios.
