This paper studies deployment of large-scale multi-agent systems (MASs), whose collective dynamics can be reduced to a PDE model under neighbor-based communication. To lower system cost, only a small number of high-cost leaders are equipped with global information, while the remaining agents rely on local interactions. In practice, the system is affected by intrinsic uncertainties, and the leaders’ global information is subject to unknown quantization. To address these issues, a Takagi-Sugeno (T-S) fuzzy approximation and an adaptive quantization-compensation scheme are developed, together with an event-triggered mechanism(ETM) to reduce global-information transmissions. Lyapunov-based analysis establishes sufficient conditions to guarantee uniform ultimate boundedness of the leader–follower closed-loop PDE system under pointwise actuation. Both simulation and real-world robot experiments include comparative reference cases, showing higher resource utilization with lower overall system cost.

Most existing control methods for nonlinear systems rely on adaptive parameters or assume full-state measurability, with limited attention paid to scenarios involving unmeasurable system states. To overcome these shortcomings, this study proposes a neuro-learning based fault-tolerant control strategy for a nonlinear two-degrees-of-freedom (2-DOF) helicopter system with sensor gain faults and an unknown dead zone. First, to address the inaccuracy of state measurements caused by sensor gain faults, a state observer is designed to reconstruct the system states, and adaptive parameters are introduced to estimate the fault in real time, providing compensation information for the controller design. A radial basis function neural network (RBFNN) is employed to address the uncertainties in the nonlinear helicopter system. In addition, the RBFNN, adaptive parameters, and bounded estimation are combined to compensate for the effects of the unknown dead zone. The stability and convergence of the closed-loop system are analyzed using the direct Lyapunov method. In simulations, the observer accurately estimates the actual system states before sensor faults occur, and it responds rapidly and reestablishes accurate state estimation when a fault is introduced. In an experimental validation using a Quanser 2-DOF helicopter platform, the proposed control method provides improved tracking performance and enhanced robustness.

To accurately compensate for system uncertainties and unknown external disturbances, this article proposes an extended observer-based integral fast terminal sliding mode control (IFTSMC) scheme for a 2-DOF helicopter system based on deterministic learning. First, during the learning phase, radial basis function neural networks (RBFNNs) are employed for dynamic modeling. By proving the satisfaction of the persistent excitation condition, RBFNNs can accurately identify the unknown dynamics, which are considered as knowledge stored in constant RBFNNs. Utilizing this acquired knowledge, a learning-based IFTSMC scheme with a learning-based ESO is proposed. The distinctive feature of this scheme lies in the designed learning-based ESO, capable of accurately observing unknown angular velocity and disturbances with low gains, thereby reducing the bandwidth requirement. Through precise disturbance compensation, the proposed controller simultaneously enhances robustness and reduces chattering. Finally, simulation and experimental results are presented to demonstrate the effectiveness of the proposed IFTSMC scheme.

Balancing and buffer allocation in mixed-model assembly lines are two critical issues that jointly determine the production efficiency of manufacturing systems. Due to the strong coupling relationship between these two problems, traditional stepwise or fragmented approaches fail to adequately capture their mutual impacts. To address this challenge, a bilevel optimization method is proposed for the joint optimization of mixed-model assembly line balancing and buffer allocation under stochastic task times and parallel workstations. Additionally, a bilevel evolutionary optimization algorithm based on deep reinforcement learning (DDRLA-SR) is designed to solve the proposed model. A trained deep reinforcement learning agent is used to generate lower-level buffer allocation responses. Furthermore, a framework for discovering deep reinforcement learning algorithms with large language models (DDRLA) is proposed. This framework automatically generates and optimizes state and reward functions in deep reinforcement learning, thereby reducing the need for expert knowledge. Experimental results show that DDRLA-SR reduces the design cost by an average of 9.9% compared with the existing method. Compared with the bilevel genetic algorithm, DDRLA-SR achieves comparable or better solution quality while reducing the average online solution time by approximately 53%. Finally, the proposed method was evaluated using two cases derived from real automotive assembly lines, further suggesting its potential value.

Dynamic modeling and motion planning of autonomous surface manipulator systems, which are novel ocean robotics comprising unmanned surface vehicles and manipulators, are crucial for their development and application. Traditional physical modeling methods have several limitations, including the requirement for extensive experiments for model parameter identification. Developing a physics-informed neural network (PINN) by integrating physical information with neural networks can address these issues. However, the accuracy of traditional PINN decreases when the nonlinear forces acting on the autonomous surface manipulator systems are unclear. To tackle these challenges, this paper proposes a serial PINN for the dynamic modeling of autonomous surface manipulator systems. This network comprises two serially connected subnetworks that predict the system’s motion states and nonlinear forces acting on it, respectively, and can be trained simultaneously. It also constructs a physical loss function based on the system’s force balance conditions to integrate physical information effectively. The accuracy of serial PINN is first validated using a two-degree-of-freedom nonlinear system, with results showing high predictive accuracy for both the system states and the nonlinear forces, and excellent generalization capabilities. Subsequently, dynamic models of 6-degree-of-freedom and 12-degree-of-freedom autonomous surface manipulator systems are developed using serial PINN, with their accuracy, computational efficiency, and generalization validated by comparisons with physical models, fully data-driven long short-term memory network, traditional PINN, and experimental results. Finally, a motion planning method based on serial PINN is proposed, which optimizes manipulator’s operation by real-time predicting autonomous surface manipulator system’s configuration changes, thus improving the entire system’s working precision.

This article proposes a singularity-free prescribed performance control (SFPPC) approach for uncertain mechanical systems to satisfy prescribed performance requirements. The approach is implemented in two stages. First, within the kinematic phase, a novel SFPPC strategy is introduced. This strategy employs a shear mapping error transformation function (SMETF) to ensure performance constraint satisfaction while circumventing the singularity issue inherent in traditional prescribed performance controllers (PPCs). Second, in the dynamic phase, an adaptive robust constraint-following controller is meticulously designed to guarantee satisfaction of all prescribed performance constraints. This controller effectively attenuates bounded uncertainties, including those that may be fast time-varying. Crucially, the controller incorporates a leakage-type adaptive mechanism to estimate the uncertainties, eliminating the need for any prior knowledge of the uncertainties. Comparative numerical simulations demonstrate that the proposed approach simultaneously guarantees singularity avoidance, prescribed performance, and robust stability.

This article investigates an elastic disturbance-compensation-based event-triggered control (EDC-ETC) problem for a communication-restricted quadrotor UAV subject to actuator saturation-backlash and external disturbances. First, to quickly and accurately estimate unknown external disturbances, cascade disturbance observers are proposed by employing multilevel disturbance error estimations; further, in order to utilize beneficial disturbances in enhancing the stability of the quadrotor UAV elastically, an EDC strategy is introduced using rigorous theoretical analysis. The effect caused by actuator saturation-backlash is removed by adopting an auxiliary system design method. Then, based on the aforementioned disturbance and constraint compensation mechanism, an EDC-ETC scheme with appropriate design parameters is developed to ensure that the trajectory tracking errors of the closed-loop system for a quadrotor UAV converge to a small neighborhood of zero while saving communication and computational resources. Finally, numerical simulations are performed to validate that the proposed scheme is feasible and better than some advanced control algorithms.

Dynamic obstacles pose one of the most significant safety challenges for autonomous robots operating in cluttered environments. This paper aims to achieve collision-free obstacle avoidance without sacrificing the prescribed performance requirements in the formation control of wheeled humanoid robots (WHRs). A novel dynamic obstacle avoidance strategy is proposed by introducing a virtual barrier potential field incorporating the obstacle’s position, velocity, and the relative angle with respect to the robot. To balance the predefined performance constraints and safety/connectivity requirements, a finite-time flexible performance function is developed to ensure finite-time convergence of the formation tracking errors to predetermined regions, while the performance boundaries can adjust flexibly when the performance constraints conflict with the collision avoidance and/or connectivity maintenance requirements. Based on the dynamic obstacle avoidance strategy and the flexible performance control, a distributed finite-time flexible performance formation control strategy is constructed, which achieves finite-time convergence of the formation tracking errors, dynamic obstacle avoidance for all robots, and connectivity maintenance among initially connected robots. Finally, the effectiveness of the proposed formation control strategy is demonstrated by both simulation and experimental results.

This article tackles the problem of achieving state consistency in a multi-agent system described by parabolic partial differential equation (PDE) through the design of a non-collocated distributed observer with event triggering. We develop a Luenberger PDE observer utilizing an event-triggered mechanism based on output errors to exponentially track the system's state. A feedback controller is introduced that relies on the estimated state. This approach ensures the consistency of the multi-agent system through boundary control and achieves convergence of the observer to the true system state. Using the Lyapunov technique, we show that the state error is uniformly bounded and the observation error is exponentially stable. Additionally, we provide a rigorous proof of the presence of a minimum time interval between triggering events.

With the rapid development of intelligent manufacturing technologies, factors such as the rapid updating of products and the complexity of humanoid robot collaborative scenarios have put forward higher requirements for the dynamic obstacle avoidance and adaptive control of industrial assembly robots. Although deep reinforcement learning offers an effective solution for robot assembly, it typically requires long training time and massive data to achieve the desired performance. However, transfer learning can enhance sample efficiency yet suffers from insufficient generalization in new scenarios. To tackle the above issues, this study proposes a pre-trained policy in the source environment to accelerate target policy training by sharing action-space features and policy network structure during transfer. Additionally, the observation space in the target environment is adjusted by concatenating robot state information with obstacle state information, capturing more critical information to address the mismatch between the observation spaces of the source and target environments. The core innovation is to first train a suboptimal policy in a simple, obstacle-free source environment for basic assembly, then transfer and fine-tune it in highly dynamic target environments involving humans and humanoid robots. This method reduces accidental collisions and enables reliable assembly in complex dynamic scenes. Experimental results demonstrate that our approach significantly improves policy learning efficiency, accelerates adaptation to new obstacle-avoidance tasks, and achieves faster convergence and stronger generalization compared with existing methods.

The control of multiagent systems can be significantly affected by collisions and uncertainties. In this article, an adaptive robust constraint-following control approach is developed for a class of nonlinear multiagent systems. Our approach consists of two phases. First, in the kinematic phase, to achieve the control goals of formation maintenance, trajectory following, and collision avoidance, a formation protocol (constraint) is constructed for each agent using local tracking errors and artificial potential functions (APF). Second, based on this protocol, to mitigate harsh uncertainties (e.g., fast time-varying), an adaptive robust constraint-following formation controller is designed in the dynamic phase. This controller not only compensates for uncertainties but also ensures that the constraint deviations are uniformly bounded and uniformly ultimately bounded. The control efficacy of the proposed approach is verified through an illustrative example.

This research addresses the fundamental challenge of achieving consensus in multi-agent systems subject to denial of-service attacks with switching network topologies. Considering that complete state measurements of the followers cannot always be obtained, to address state unavailability, Luenberger-based observers are designed for estimation. Boundary control and a dynamic event-triggered mechanism are designed based on the observations to alleviate the shortage of communication resources. The proposed control scheme accounts for the full duration of system abnormality, including the post-attack interval preceding the next event trigger, and is proven to guarantee consensus. Based on the parameters of the controller, the average dwell time of the switching topology, the attack duration and the attack frequency, the sufficient conditions for the asymptotic consensus of the system are given. Theoretical analysis demonstrates that the proposed control mechanism can effectively achieve secure consensus. In addition, it is shown that the designed triggering mechanism can exclude Zeno behavior. Numerical simulation studies are conducted to verify the effectiveness of the proposed control algorithm.

This study focuses on the reinforcement learning (RL)-based consensus tracking control of nonlinear multiagent robot systems (MARSs) with event triggering mechanism. Each agent of the MARSs is composed of a three-link rigid robot and a flexible payload, which can be assumed to be a Eulbernoulli beam. Based on the assumed mode method (AMM), the infinite distributed parameter model of the robot-payload system is approximated as a finite dimension model, and the dynamic performance of the robot system is controlled with the use of boundary control input. First, a RL control strategy based on actor-critic structure is adopted to maintain the consensus angles tracking of all agents while suppress the load vibration. Second, considering the communication bandwidth problem in practical applications, an event-triggered mechanism is utilized to reduce the transmission burden based on relative threshold strategy. Furthermore, the semi-global uniformly ultimately bounded (SGUUB) property of the closed-loop system is derived to guarantee the state errors can converge to the small neighborhoods of the origin. Finally, the effectiveness of the proposed control strategy is demonstrated by numerical simulations.

This paper investigates the fixed-time adaptive containment control problem for Distributed Parameter MultiAgent Systems (DP-MASs) subject to unknown nonlinear uncertainties. Addressing the complexity of modeling largescale swarms, a macroscopic dynamic model of the system is established using Partial Differential Equations (PDEs) to realize spatially continuous trajectory planning. To address the internal unknown nonlinear dynamics and external disturbances, a Fuzzy Wavelet Neural Network (FWNN) approximation strategy is proposed. Based on Lyapunov analysis, the stability of the closed-loop system is verified. Numerical simulations demonstrate that the system achieves stabilization before the preset time, validating the effectiveness and feasibility of the algorithm.

In this paper, a fixed-time adaptive deferred constrained control strategy is proposed for a flexible single-link manipulator system with input saturation. Fuzzy Neural Networks are utilized to estimate the unknown dynamics of the flexible manipulator system as well as the errors caused by input saturation. To address output constraints imposed within a prescribed time period, a time-shift function and an adjusted barrier function are introduced. The system’s stability is rigorously proven using the direct Lyapunov method. Finally, numerical simulations and experimental results are presented to validate the effectiveness and superiority of the proposed control approach.

This paper investigates the adaptive fixed-time containment control problem for distributed parameter multi-agent systems (DP-MASs) under input quantization. To address the deployment issue of large-scale multi-agent systems, a partial differential equation (PDE) is employed to establish the macroscopic dynamic model, by which the coordination task is reformulated as a dynamic target tracking problem. In practical scenarios, a hysteretic quantizer is integrated into the control channel to effectively alleviate the communication burden. Moreover, to compensate for the effects of inherent system uncertainties and quantization errors, this paper presents a control scheme combining radial basis function (RBF) neural network approximation and adaptive quantization compensation. Based on Lyapunov stability analysis, it is proved that the tracking error converges to a small neighborhood of the origin within a fixed time, and a series of numerical simulations are conducted to validate the effectiveness and feasibility of the proposed control strategy.

This article proposes a continual deterministic learning (DL) control framework for unknown nonlinear systems to achieve efficient knowledge accumulation and high-performance control over sequential trajectory tracking tasks. First, an adaptive incremental node generation mechanism is developed to enable the neural network (NN) to dynamically expand its structure along evolving input trajectories, thereby maintaining sufficient approximation capability under varying operating conditions. Based on DL theory, the local closed-loop dynamics along each trajectory are accurately identified under the persistent excitation condition (PEC) and stored in a constant NN representation as reusable dynamical knowledge. By continuously accumulating and integrating the learned knowledge from sequential tracking tasks, a unified multitrajectory dynamical representation is constructed, enabling rapid adaptation to new trajectories. Furthermore, to mitigate performance degradation caused by dynamical discrepancies during sequential learning, an enhanced node mechanism is introduced, resulting in the development of an experience-enhanced DL controller (EEDLC). Simulation and experimental results on a helicopter platform demonstrate that the proposed approach significantly improves learning efficiency across sequential tasks while maintaining satisfactory closed-loop tracking performance.

Stable grasping operations play a pivotal role in enabling robotic dexterity. To achieve a level of proficiency akin to humans, robots must effectively integrate visual and tactile information throughout their grasping tasks. However, relying solely on visual information may not enable robots to achieve dexterous operations. To address this issue, this study explores the application of visual–tactile fusion in robot grasping operations and proposes the YGC model. Specifically, we decompose the grasping operation into three subtasks: target detection, pose generation, and grasping state detection. Firstly, in order to enhance the feature extraction capabilities of the YOLOv5s backbone network, we propose the integration of a multi-scale feature fusion (MSFF) module, which replaces the existing C3 module. In pose generation, we compare various input channels to identify the optimal input method. For grasping state detection, we evaluate the performance of single-modality and multi-modality inputs. Secondly, experimental results show that by replacing the MSFF module, the mean average precision (mAP) value improves by 3.59

To address the significant errors in traditional odometry-based estimation caused by posture variations when wheel-legged robots traverse complex terrains, this paper formulates terrain parameter estimation as a nonlinear regression problem based on time series and proposes a proprioception-based learning approach. Specifically, a state-space modeling framework is employed, introducing the Mamba model to capture the implicit mapping between multi-source robot states and terrain parameters. Its efficient long-sequence modeling capability allows accurate representation of temporal dependencies and nonlinear couplings in complex dynamic processes. To validate the method, LSTM and CNN-Attention-LSTM models are selected as comparative baselines for experimental analysis. The proposed approach does not rely on accurate physical models and can learn the complex mapping between robot states and terrain parameters directly from data, providing a novel data-driven solution for terrain perception of wheel-legged robots in complex environments.

Ensuring the safe and reliable operation of quadrotor unmanned aerial vehicles (UAVs) under actuator faults is critical for practical applications. Conventional control methods often face challenges to simultaneously accommodate system constraints and adapt to actuator faults, potentially degrading performance under uncertainty. To address these limitations, this paper proposes an adaptive fault-tolerant control framework integrating Model Predictive Control (MPC) with Deep Deterministic Policy Gradient (DDPG) reinforcement learning. Within this framework, MPC serves as the baseline controller by generating feasible inputs that satisfy constraints. Key to this approach is the introduction of a Lyapunov-based contraction constraint, which replaces conventional terminal constraints to guarantee system stability and recursive feasibility while significantly reducing the required prediction horizon and computational burden. Concurrently, the DDPG agent learns to generate compensatory and auxiliary control signals, thereby mitigating external disturbances and actuator faults. A key innovation lies in employing a parameterized NMPC formulation as the Actor within the DDPG framework. This enables online adaptation of cost function weights and fault compensation parameters through policy optimization. The simulation results demonstrate that our hybrid MPC-DDPG approach achieves accelerated trajectory recovery and exhibits superior tracking accuracy, fault tolerance, and robustness under actuator fault scenarios.