ABSTRACT Central pattern generators (CPGs) provide a powerful biological principle for rhythmic coordination, but artificial CPGs often depend on manually designed coupling functions and topology‐specific analysis. This proposal focuses on the generalisation mechanism of a swarm‐inspired central pattern generator (SCPG), in which each oscillator learns local attention‐based interactions with its neighbours and collective synchronisation emerges from these distributed decisions. The central hypothesis is that learnt local attention can serve as a general coupling rule that transfers across network scale and target synchronisation mode. To test this hypothesis, the proposal centres on scale/task generalisation experiments, a simplified phase‐model experiment that probes whether a fixed phase‐dependent coupling rule can generalise across network sizes and ablation studies that identify the architectural source of generalisation. The expected outcome is a focused mechanistic account of why SCPG can replace handcrafted CPG coupling rules with scalable learnt local interaction.
ABSTRACT This work reviews efforts to develop systems for in‐motion wireless power delivery, that is, the transmission of energy to moving receivers (RX) through untethered interaction. The application in focus is supplying small‐to‐medium‐sized grounded mobile robots, but the insights extend beyond this context and the review serves as a platform for developing such systems. A review of the literature strongly indicates that such systems must rely on inductive coupling with magnetic resonance (MR), a technique that introduces its own set of technical challenges in real‐world scenarios. The most significant of these are analysed, and corresponding solutions from the literature are presented and critically reviewed. This review differs from earlier work in that prior findings are examined in the context of multi‐transmitter (TX) multi‐receiver systems, as this configuration was found to be preferable for sufficiently efficient dynamic wireless power transfer.
ABSTRACT To address the challenges of strong nonlinearity, high thermal inertia and severe data discontinuity caused by ‘off‐peak’ storage strategies in large‐capacity thermal storage systems within cigarette manufacturing, this paper proposes a physics‐enhanced long short‐term memory (PE‐LSTM) modelling framework. By systematically integrating engineering thermophysics with deep learning, the framework enhances model credibility and generalisation across multiple layers. Specifically, a temporal causality integrity check algorithm is developed to mitigate intermittent sampling issues, preventing the learning of spurious dynamic responses across data gaps by blocking gradient backpropagation during non‐operational periods. Furthermore, physics‐aware feature engineering reconstructs highly coupled variables, such as heating power and coefficient of performance (COP), based on energy conservation laws. A multi‐task collaborative learning architecture is then employed, utilising an asymmetric weighted loss function to explicitly embed energy consistency residuals and reverse Carnot cycle constraints. Compared with task‐specific single‐output standard LSTM baselines, the proposed PE‐LSTM maintains comparable test‐set accuracy for tank temperature, heat‐pump power, COP and stepwise heat production within a unified multi‐output framework. These results indicate that the proposed model provides a physically consistent modelling foundation for energy analysis and operation optimisation in tobacco industrial thermal systems.
ABSTRACT In the pursuit of proactive human–robot collaboration, mutual cognition depends not only on smarter robots but also on more capable human operators able to anticipate robot actions through enhanced cognitive function. In this study, we propose a digital twin (DT)‐based mixed reality system that delivers immersive and real‐time representations of robot states, task progression and motion trajectories which establish the system‐level preconditions for human prediction and informed intervention during collaboration. The framework is validated through system‐level properties—spatial fidelity, temporal relevance and robustness. These constitute necessary technical conditions for improved cognitive function, such as trust and anticipation in proactive human–robot collaboration, rather than directly measuring human cognitive outcomes. Validation involves a full‐factorial case study that assesses positional accuracy and communication delay across varying motion trajectories, tool centre point speeds and network conditions. Results demonstrate positional precision of 0.105, 0.150 and 0.158 mm in , and axes, respectively, under optimal conditions, a high correlation between actual and virtual locations ( > 0.97) and a communication latency of 54.85 ms and DT rendering latency of 7.09 ms. This contributes to proactive human–robot collaboration research by illustrating system‐level technical preconditions necessary to support human‐cognitive function such as trust and anticipatory control for scalable human–robot collaboration in industrial applications.
ABSTRACT With the rapid proliferation of unmanned aerial vehicles (UAVs) in the military, logistics and public safety sectors, the threat of illegal intrusion, information theft, and security breaches has increased. This paper provides a comprehensive review of recent developments in counter‐unmanned aerial vehicle (counter‐UAV) systems, focussing on detection and defence technologies. At the detection level, we examine multimodal sensing technologies, including radar, infrared, optical, acoustic and RF systems, and discuss their advantages and limitations. At the defence level, we explore active countermeasures such as electronic jamming, net capture, lasers, directed energy and swarm cooperative interception. A comparative analysis of these methods is provided, highlighting their effectiveness in different operational scenarios. Furthermore, the paper discusses the development trends of multi‐source information fusion, intelligent decision‐making and autonomous defence technologies. As counter‐UAV systems evolve, intelligent, collaborative and distributed fusion approaches are expected to play a pivotal role in enhancing the overall system performance. This paper aims to offer a systematic reference for researchers and practitioners involved in the design and implementation of low‐altitude security systems.
ABSTRACT Multiagent reinforcement learning (MARL) has emerged as a key enabling technology for intelligent coordination in cyber‐physical robotic systems, such as automated warehouses. However, the effectiveness of MARL in real‐world scenarios is often hindered by challenges, such as sparse reward signals and partial observability, which limit the agents' ability to learn optimal behaviours. In this paper, we propose a novel method to design MARL for warehouse management and implement it within a game engine to facilitate real‐world visualisation for industry users. The contributions of this paper are twofold. First, we introduce temporal attention‐enhanced counterfactual multiagent reinforcement learning (TA‐COMA), a novel method designed to enhance performance in environments with sparse rewards and partial observations by integrating temporal advantages. Second, we propose a new approach to define intermediate rewards in sparse reward environments, exemplified through a warehouse scenario. We also demonstrate, for the first time, the implementation of a multiagent reinforcement learning strategy for warehouse environments within the Unity game engine, utilising the ML‐Agents toolkit. The simulation results demonstrate that intermittent rewards significantly improve learning performance. TA‐COMA consistently achieves higher asymptotic mean cumulative rewards than COMA, PPO and shared experience actor–critic (SEAC). We further show that the variance of the temporal advantage in TA‐COMA is reduced by approximately compared to COMA, improving temporal coherence whilst preserving informative advantage structure and stabilising learning. The Unity platform further enables the creation of a realistic 3D warehouse environment and immersive visualisation of multirobot behaviour using a virtual reality (VR) headset.
ABSTRACT To address the poor performance of navigation line extraction for agricultural robots in complex open‐field corn ( Zea mays L.) environments during the 6–8 leaf growth stage, this study proposes an accurate and efficient navigation baseline extraction method based on an improved YOLO11n model, specifically designed for complex corn field conditions. First, an improved corn root–stalk detection model, termed RO‐YOLO, is proposed (YOLO is short for You Only Look Once). By incorporating the EBlock, a low‐resolution self‐attention module and a multifeature fusion module into the YOLO11n framework, the detection accuracy of the model is significantly enhanced. Next, corn root–stalk positions are precisely localised based on the bounding boxes generated by the model, and representative feature points of corn crop rows are extracted. Subsequently, a feature point screening strategy based on coarse lateral position partitioning and precise longitudinal near‐field filtering is proposed to select valid feature points, and the navigation baseline is fitted using the random sample consensus (RANSAC) algorithm. Finally, the navigation line is extracted by solving the angle bisector. Experimental results demonstrate that the improved RO‐YOLO achieves a precision, recall and AP@0.5 (average precision at an intersection over union [IoU] threshold of 0.5) of 91.4%, 88.3% and 92.7%, respectively, for corn root–stalk detection. Moreover, the proposed navigation line extraction algorithm for the corn 6–8 leaf stage attains an average fitting time of only 51.6 ms and an accuracy of up to 92.0%, ensuring both high precision and real‐time performance of navigation line extraction.
The use of chemical treatments for weed control is increasingly challenged by weed resistance, regulatory restrictions on chemicals and the risk of soil and water contamination that can pose health hazards. Chemical weed treatments are unsustainable, prompting exploration of alternative methods such as boiling water, electrocution and directed fire. However, these approaches are limited: water-based treatments require a reliable water supply in the field, and electric or fire-based methods pose additional environmental risks. This work proposes a targeted light amplified stimulation of emission by radiation (LASER) treatment and selective spraying system integrated with automatic weed identification. The system, mounted on the rear of a tractor, utilises a bispectral imaging setup to distinguish weeds from crops. Close-to-crop weeds are automatically selected for LASER treatment, whereas a selective sprayer targets other weeds with a glyphosate globule. This integrated system approach is named 'Hyperweeding'. The results demonstrate successful separation of row crops from weeds, achieving a contamination-free crop with a significant reduction in glyphosate usage, and effectively treating weeds at speeds of 0.1 m.
This work promotes traditional pair-wise interaction (PWI) to higher-order interaction (HOI) models in real robot swarms, aiming to faithfully represent complex phenomena such as the agile and stable turning manoeuvres observed in natural bird flocks. However, existing HOI models are typically based on idealised particle-like assumptions, which are rarely applicable to real robot swarms due to physical and operational constraints. To address this gap, we propose an embodied framework for extending PWI to HOI in real robot swarms. Our framework explicitly accounts for physical dimensions and operational constraints by incorporating suitable potential functions and designing a velocity coordination component consistent with HOI principles. The proposed approach is evaluated through comprehensive simulations and physical experiments under two distinct scenarios: collective evasion and trajectory tracking. Results show that the extension to HOI models contributes to improvements of up to in overall performance, in responsiveness and in cohesion. Beyond deployment of HOI models in real robot swarms, our work paves a practical way for using robot swarms to investigate HOI topologies that underlie collective behaviours, such as bird flocking, fish schooling and mammal herding.
Climbing robots have broad potential for various practical applications. In this study, a five-degree-of-freedom inchworm robotic platform is developed using modular T/I joints and its kinematic model is derived based on the Denavit-Hartenberg approach. Multiple gait modes, including creeping, rotation, flipping and obstacle negotiation, are investigated. To address the high energy consumption during the creeping gait, an optimisation framework constrained by both kinematic and dynamic factors is established. This framework adopts a deviation-tracking strategy that combines cubic polynomials in the Cartesian space with 7th-order polynomials in the joint space. Mathematical models for energy evaluation, trajectory formulation and spatial curve similarity are also developed. Additionally, an improved adaptive particle swarm optimisation algorithm is proposed, integrating three enhancements: a Sigmoid-based update rule, a position-updating scheme inspired by the golden sine algorithm and a tabu-region constraint mechanism. These modifications result in faster convergence and higher precision. Finally, a complete control system is designed, and experimental tests conducted under the creeping and rotating gaits verify the practicality and effectiveness of the proposed motion-planning method.
The inspection of long-range tunnels has become an important challenge in underwater robotics, as it involves a trade-off among coverage, efficiency, and an energy-related cost metric. In this paper, we propose a space-aware trajectory planning algorithm to generate high-coverage inspection trajectories for long-range tunnel inspection. The proposed method has distinctive features in the following aspects: (1) three-dimensional modelling and initial trajectory generation of tunnel structure based on spatial perception; (2) an adaptive spiral trajectory model in cylindrical coordinates to alleviate blind spot problems near tunnel ends; (3) a multi-objective optimisation strategy based on the non-dominated sorting genetic algorithm II (NSGA-II) to balance coverage effectiveness, trajectory smoothness, and energy-related cost. Simulations in representative tunnel environments demonstrate that the proposed method achieves high coverage rates while substantially reducing redundant motion. These results highlight both the theoretical and practical value of the framework for intelligent tunnel inspection, with promising implications for enhancing the safety of large-scale water infrastructure.
Unmanned aerial vehicle (UAV) swarms hold significant promise for scalable environment exploration, cooperative transportation and urban air mobility. Nevertheless, trajectory planning for swarms remains challenging due to high computational complexity, stringent spatiotemporal constraints and limited attention to energy efficiency. In this work, we present a unified trajectory planning framework that simultaneously promotes energy efficiency and ensures spatiotemporal consistency. Building upon the minimum control (MINCO) trajectory representation, we extend the differentially flat parameterisation from single-UAV settings to multiagent systems, enabling efficient representation of swarm trajectories through sparse spatiotemporal parameters. Differentiable cost functions are designed to explicitly capture synchronised arrivals, temporal logic of tasks and collision-avoidance constraints, whereas analytical gradients are derived to facilitate scalable optimisation. Extensive simulations with swarms of up to 15 UAVs demonstrate the scalability of our method in generating smooth and dynamically feasible trajectories. Compared to state-of-the-art baselines, the proposed framework significantly improves the interagent safety margin (by 35%) and achieves substantial energy efficiency, reducing physical energy consumption by over 10% and control effort by up to 22%.
Dissolved oxygen (DO) is a vital parameter in aquaculture, directly influencing fish health, growth and overall productivity. However, existing DO prediction methods largely rely on complex physical models that demand numerous parameter measurements and high computational cost. This study introduces a hybrid framework that integrates transfer learning (TL) with physics-guided neural networks (PGNNs) based on long short-term memory models (PG-LSTM-TL), to accurately predict pond DO levels from only 1 month of sparse and irregularly sampled nighttime data. The TL strategy leverages knowledge from a data-rich tidal estuary and adapts it to a data-scarce aquaculture setting, whereas the physics-guided component embeds two key domain constraints: The temperature-dependent DO solubility governed by Henry's law and the monotonic nighttime DO depletion trend. Unlike conventional TL approaches that primarily rely on fine-tuning or statistical feature alignment, we embed physics-guided constraints directly into the model learning stage. This integration mitigates negative transfer arising from environmental heterogeneity while preserving physically plausible multi-step DO trajectories. Results show that, for each daily cycle (9:00 p.m.-5:00 a.m.), the model uses just the first four hourly measurements to recursively forecast DO until dawn, achieving an overall mean absolute error (MAE) of 0.3593 mg/L and root mean square error (RMSE) of 0.4813 mg/L. This represents improvements of 35.5% (MAE) and 32.4% (RMSE) over conventional models. By combining physical insights with data efficiency, the proposed approach offers a practical and scalable solution for reliable DO prediction in small-scale aquaculture systems worldwide.
Deployment of autonomous driving trucks (ADTs) in open-pit mining enables more flexible transport operations than human-driven trucks; but this flexibility significantly expands the scheduling search space and makes it more difficult to find near-optimal solutions on large scales. To address this challenge, this study proposes a bi-layer divide-and-conquer framework for large-scale ADT scheduling. In the upper layer, a mixed-integer linear programming model is formulated to determine the minimum fleet size (FS) required to satisfy production targets within a shift, thereby reducing the dimensionality of the scheduling problem. Given the optimised FS, the lower layer solves the scheduling problem using a reinforcement learning (RL)-assisted evolutionary programming approach. Specifically, a deep Q-network is embedded in a genetic algorithm to adaptively adjust the crossover and mutation probabilities, improving search efficiency and enhancing the algorithm's ability to escape local optima. Experiments based on real-world mining scenarios in Inner Mongolia including over-100-truck scenarios show that the proposed framework can reduce FS by more than 10% and fuel consumption by over 20% compared to current methods. Overall, the framework reduces problem dimensionality without compromising optimality and supports efficient scheduling for large-scale mining operations.
Topology-driven model predictive control (T-MPC) has demonstrated significant performance in navigating dynamic environments by planning trajectories in distinct homotopy classes in parallel. However, its performance is limited by deterministic assumptions about obstacle motions and a cost function that ignores social norms. This paper presents a unified framework, socially compliant topology-driven model predictive control under uncertainty (SCU-T-MPC) that addresses these limitations. First, we introduce the concept of probabilistic homotopy classes by integrating multi-modal uncertainty into the topological guidance planner, enabling robustness to various potential futures. Second, we formulate a risk-aware parallel optimisation scheme where each local planner incorporates chance constraints and a risk-adjusted cost. Finally, we learn a social preference model from human demonstration data that map topological features to a social cost, biasing the trajectory selection towards socially compliant behaviours. We provide theoretical guarantees on the robustness of our approach. Extensive simulation experiments with a mobile robot navigating among pedestrians show that SCU-T-MPC outperforms existing planners in terms of safety, efficiency and social acceptability.
Autonomous parking remains a challenging task due to the need for accurate trajectory tracking, smooth steering, and stable heading control under diverse manoeuvring conditions. Conventional model predictive control (MPC) can handle system constraints effectively, but its performance depends heavily on manually tuned cost weights. This paper proposes a reinforcement learning-assisted MPC (RL-assisted MPC) framework to improve autonomous vehicle parking performance. A deep Q-network (DQN) agent is trained to dynamically select the cost function weights of an MPC controller, enabling real-time adaptation based on the vehicle's current state. The hybrid approach leverages the predictive optimisation capability of MPC together with the adaptive decision-making of RL, enabling the controller to adjust trade-offs in real time without manual re-tuning. The framework is evaluated across five different parking scenarios and compared against static-weight MPC baselines. Experimental evaluations demonstrate that the proposed RL-assisted MPC framework achieves comparable or better lateral tracking accuracy, while consistently providing smoother steering behaviour and improved heading stability compared with baseline controllers using static MPC weights. The proposed framework is further evaluated under randomly selected and previously unseen initial vehicle positions, demonstrating its robustness and generalisation across diverse parking configurations. The results demonstrate that RL-assisted MPC improves robustness and generalisation in automated parking systems, highlighting the potential of combining model-based predictive control with RL for autonomous driving.
High-precision trajectory tracking control of space flexible manipulator represents a significant research focus of contemporary research and poses great challenges in both academia and engineering. To address the issue of low control precision in space flexible manipulator, which arises from highly nonlinear dynamics in complex spacecraft environments, the LuGre friction model is incorporated into the dynamic equation to improve the accuracy of frictional dynamic behaviour modelling. Subsequently, a reinforcement learning-based sliding mode control (RL-SMC) method is developed to achieve precise approximation and compensation of uncertain nonlinearities within the space flexible manipulator system. The employed RL framework is based on the actor-critic architecture, where the actor neural network generates the control policy, whereas the critic neural network evaluates the policy and continuously provides feedback regarding the system state. This control method uses a radial basis function neural network (RBFNN) combined with the SMC to minimise approximation error. In complex space environments, the actor-critic framework enhances the approximation of nonlinear dynamics for a space flexible manipulator and facilitates more efficient adaptation to variations in system dynamics. In addition, joint angle output constraints are implemented to manage the restricted motion of the space flexible manipulator in confined workspaces, aiming to prevent collisions during operation and avoid structural damage. Finally, the stability of the closed-loop system is rigorously established using the Lyapunov stability theory. Numerical simulations demonstrate the efficacy of the proposed approach in improving both control precision and environmental adaptability of the space manipulator.
In typical operational scenarios such as turbine blade inspection, robots are subject to complex spatial constraints and are prone to mechanism-environment interference, which restricts their motion and may even render them inoperative. To address this issue, this paper proposes a novel 10-degree-of-freedom (10-DOF) continuum robot configuration and, on this basis, develops a bounded nonlinear least-squares (NLS) inverse kinematics (IK) framework for precise motion control under joint limits and tube-shaped workspace constraints. Specifically, a tube-aware rapidly-exploring random tree connect (RRT-Connect) planner is first employed to compute a coarse joint-space path with edge-wise feasibility checking; then, sequential least squares programming (SLSQP) refines it into a smooth skeleton; finally, follow-the-leader (FTL) performs dense Cartesian micro-stepping via bounded NLS, while enforcing segmented virtual-tube soft constraints along link-sampled points. Simulation results in CoppeliaSim demonstrate that the proposed method generates smooth and safety-compliant trajectories in confined environments and robustly tracks the target blade edge curve, effectively mitigating branch jumping and orientation discontinuities. Quantitative metrics, including tracking root-mean-square error (RMSE), tube-margin, violation statistics, and per-step computation time, indicate favourable feasibility and stability. Overall, this work provides an effective joint solution for global-local trajectory planning and constrained IK of redundant continuum robots in strongly constrained cavity inspection tasks.
Reduced-order models (ROMs) are widely employed in biped robot control due to their computational efficiency, but their simplified representations often neglect critical nonlinear dynamics, leading to limited robustness under real-world disturbances. To overcome this limitation, this paper introduces a robust hierarchical control framework that explicitly compensates for unmodelled dynamics and provides theoretical stability guarantees. The proposed architecture consists of two layers. At the high level, a hybrid-linear inverted pendulum (HLIP) model generates real-time gait commands, whereas an feedback law accounts for dynamic uncertainties introduced by model simplification. A Lyapunov-based analysis is used to rigorously establish the stability of each planned foothold. At the low level, a whole-body controller tracks both swing-leg and centre-of-mass trajectories by solving a quadratic programme that maps task-space accelerations to joint torques. The framework is validated in simulation and hardware experiments on the BRUCE platform. On flat ground, BRUCE maintains a steady walking speed of 0.3 m/s. When confronted with uneven terrain-simulated by randomly distributed 2.5-cm planks as unmodelled disturbances-the robot preserves balance and velocity tracking. Comparative evaluations against the divergent component of motion (DCM) and reinforcement learning (RL)-based methods demonstrate superior velocity tracking performance of the proposed approach, confirming its ability to reconcile the computational tractability of ROMs with the robustness missing in many traditional and learning-based controllers.
With the increasing demand for automated inspection solutions in complex industrial environments, existing robotic platforms face significant limitations in terms of endurance, payload capacity and obstacle-crossing capabilities. In this paper, we present a novel inspection robot system based on the CubeTrack tracked platform, featuring a large configuration space achieved through the integration of a manipulator and advanced mobility mechanisms. Our system incorporates a quad-slider elliptical trammel mechanism (Qs-ETM) that enables geometry-changing tracks for enhanced terrain adaptability while maintaining track tension stability. To address multi-layer navigation challenges, we propose an efficient trajectory planning algorithm that extracts traversable planes from three-dimensional (3D) point clouds and constructs a lightweight plane graph for path optimisation. Additionally, we develop a flipper control algorithm that uses only low-cost local sensor measurement (time-of-flight [TOF] sensors and inertial measurement unit [IMU]) to enable autonomous stair navigation without pre-mapped environments. The inspection system integrates multiple sensors, including light detection and ranging (LiDAR) sensor, RGB cameras, gas sensors and thermal cameras, providing comprehensive monitoring capabilities for industrial inspection demands. Extensive real-world experiments demonstrate the system's effectiveness in navigating complex environments with stairs, multiple layers and narrow passages, validating both the mechanical design and algorithmic approaches for practical industrial inspection tasks.
