The integration of robotics in agriculture has gained significant attention due to its potential to revolutionize farming practices. This paper presents a state-of-the-art review on the application of robotics in agriculture, focusing on research challenges and future directions. The review covers various aspects of agricultural robotics, including crop monitoring, precision farming, harvesting, and livestock management. It provides an overview of the current state of agricultural robotics, highlighting key advancements and achievements. The paper identifies and analyzes the research challenges involved in implementing robotics in agriculture, encompassing technical issues, economic factors, and regulatory considerations. Moreover, it explores future directions by examining emerging technologies like artificial intelligence, machine learning, and their integration with Internet of Things and big data analytics. The paper emphasizes the importance of interdisciplinary collaboration to address the challenges and drive the development of standardized protocols, robust algorithms, and cost-effective solutions. This state-of-the-art review serves as a valuable resource for researchers, practitioners, and policymakers, offering insights to advance the field of agricultural robotics and promote sustainable and efficient farming practices.
The soft ankle exoskeleton, as a wearable assistive device, provides active assistance to the ankle joint, thereby enhancing walking performance and reducing the metabolic cost of the wearer. In this study, a multi-layer adaptive control strategy integrating decision and execution layers is proposed to enable personalized ankle assistance. At the decision layer, a gait recognition model is developed to identify the wearer’s gait in real time and to generate corresponding assistance profiles based on the recognition results and the desired biological ankle joint torque. In addition, an enhanced finite state machine is incorporated to accurately predict assistance timing across gait phases. At the execution layer, an iterative adaptive control scheme combining Bayesian optimization with PID control is employed to achieve real-time tracking and error compensation of the desired assistance profiles, enabling dynamic self-adaptation of control parameters. Experimental results demonstrate that the proposed multi-layer adaptive control strategy exhibits strong robustness in force tracking, delivering peak assistance forces of up to 120 N with peak tracking errors maintained within 5
Robotic peg-in-hole insertion is a fundamental operation in industrial assembly; however, its automation remains challenging because of tight tolerances, variations in component geometries, contact uncertainties, and the complexity of real-world operating environments. Conventional preprogrammed robotic systems often struggle to cope with such uncertainties, motivating the adoption of reinforcement learning (RL) as a more adaptive paradigm for robotic insertion tasks. This review provides a comprehensive analysis of RL-based approaches for peg-in-hole insertion and examines their evolution from traditional reinforcement learning algorithms to modern deep reinforcement learning (DRL) frameworks. This review systematically categorizes the literature according to the learning paradigm, sensing modality, control strategy, and training environment. Key methodologies are discussed, including Q-learning, SARSA, deep reinforcement learning, model-based RL, meta-RL, transformer-based reinforcement learning, transfer learning, and domain randomization techniques. Particular attention is given to multimodal perception strategies that integrate visual, force, and tactile information, as well as hybrid learning frameworks that combine RL with imitation learning and classical control methods to improve the robustness, training efficiency, and deployment reliability. In addition, this review analyzes recent advances in sim-to-real transfer, digital twin technologies, and real-time deployment strategies that facilitate the transition of learned policies from simulations to industrial robotic systems. A comparative analysis of representative studies is presented to highlight current trends, methodological differences, and practical limitations in the literature. Despite substantial progress, several challenges continue to limit large-scale industrial adoption, including simulation-to-real transfer, sample efficiency, safety, generalization, interpretability, and deployment scalability. Finally, emerging research directions are discussed, including multimodal learning, digital twin-assisted training, hybrid control architectures, multi-agent reinforcement learning, and real-time edge deployment. By consolidating recent advances, limitations, and future opportunities, this review provides a structured reference for researchers and practitioners seeking to develop next-generation intelligent robotic assembly systems.
Magnetic adsorption inspection robots have great potential applications due to their unique characteristics, such as large adhesive force and low surface cleanliness requirements. However, current magnetic adsorption inspection robots generally suffer from issues such as low obstacle-surmounting efficiency, inadequate obstacle-surmounting height, and insufficient stability during the obstacle-surmounting process. By studying wheeled and legged robots, a dual (passive and active) obstacle-surmounting mode magnetic adsorption robot and its gait design are proposed. Additionally, a magnetic dynamic adjustment (MDA) mechanism is proposed to improve the obstacle-surmounting height of the robot in the passive obstacle-surmounting mode (POSM) and the robot’s stability in the active obstacle-surmounting mode (AOSM), thereby significantly enhancing its obstacle-surmounting capability. Mechanical analyses are conducted for the obstacle-surmounting process of the robot in the passive and active modes, utilizing the moment equilibrium method and center of gravity projection method, respectively. MDA schemes in the POSM and AOSM are derived. Subsequently, a kinematics simulation is carried out to optimize the magnetic dynamic regulation scheme. Finally, experiments are conducted to verify the robot’s performance. The experimental results demonstrate that, in the passive obstacle-surmounting mode, the proposed MDA mechanism increases the maximum traversable obstacle height from 17 mm to 30 mm. In the active obstacle-surmounting mode, the robot can stably negotiate 70 mm vertical protrusions, significantly improving both the success rate and the dynamic stability of the robot during the crossing process.
Existing voice-controlled home service robots are either prohibitively expensive (INR 150,000–250,000), rely on fragmented subsystems that address only navigation or IoT control but not both, or are restricted to English-only interaction. None of the reviewed systems provide an affordable, unified, multilingual, and mobile home assistant suitable for linguistically diverse households. This paper presents Harley, a novel experimental integration of autonomous indoor navigation, MQTT-based smart home control, and multilingual conversational AI into a single low-cost robotic platform (under INR 10,000, at approximately 19 ^2 , four rooms) demonstrated high voice recognition accuracy, improved navigation success rate, and reliable IoT device control with low latency. Additionally, the system achieved an average conversational quality rating of 4.1/5.0 across multiple languages. These results establish that high-performance home automation robotics addressing the five limitations documented in Table 2 is achievable at a fraction of current market cost, with potential relevance to multilingual and cost-sensitive environments, including those serving elderly users in developing economies.
Autonomous navigation in dynamic environments requires robots to make effective decisions to avoid dynamic obstacles and reach the destination. Traditional navigation methods rely on manually designed rules that cannot readily adapt to environmental changes. Although reinforcement learning (RL) has demonstrated superior decision-making capability, it still suffers from slow training convergence and limited navigation safety. To overcome these limitations, this paper presents an RL-based autonomous navigation method based on the twin delayed deep deterministic policy gradient (TD3) algorithm. A curiosity-driven reward mechanism is designed to encourage exploration of unknown states, thereby accelerating convergence and improving training efficiency. Furthermore, an artificial potential field (APF)-driven mechanism is integrated to provide safety guidance by computing the virtual resultant force, thereby guiding the robot to learn obstacle-avoidance behaviors and enhancing navigation safety. Simulation results demonstrate that the proposed method outperforms the baseline methods in terms of navigation success rate and training convergence in dynamic environments, while real-world experiments validate its effectiveness in practical deployment.
Space manipulator play an irreplaceable role in in-orbit tasks such as routine maintenance of large space structures. To increase success rates and improve robustness of mission execution, the initial configuration of the manipulator approaching the target not only needs to exhibit excellent dexterity, but also minimize joint position changes after the end-effector travels a certain distance. To select a suitable set of joint angles from multiple inverse solutions of a given pose, a strategy for selecting the initial configuration is proposed, which uses normalized manipulability and configuration closeness as performance criteria. To improve the solution efficiency of inverse kinematics, the joint optimization range of the manipulator is obtained based on the value of the objective function, which combines normalized manipulability and configuration closeness. Due to its advantages, such as simple parameter setting and good convergence, the Particle Swarm Optimization (PSO) algorithm is employed to solve the configuration selection problem. Finally, simulation results are presented and demonstrate the effectiveness of the proposed configuration selection strategy.
Tactile sensing is critical for restoring haptic feedback in Minimally Invasive Surgery (MIS), where the surgeon’s natural sense of touch is diminished. Advances in Micro-Electro-Mechanical Systems (MEMS), piezoelectric materials, photonic sensing, and artificial intelligence (AI) have enabled compact, high-resolution tactile interfaces capable of measuring interaction forces up to 1.2 N with ≈1.2
Human-machine interfaces for robotic platforms have traditionally relied on joysticks, keyboards, and teleoperation hardware that impose cognitive and physical barriers, particularly for users with disabilities. This paper presents the design, implementation, and experimental evaluation of a fully integrated, Internet of Things-enabled gesture-controlled robotic system comprising a four-degree-of-freedom (4-DOF) robotic arm and an omnidirectional ground vehicle. The system employs an Motion Processing Unit – 6-axis (6050 series) (MPU6050) six-axis inertial measurement unit and a five-channel flex sensor array embedded within an instrumented glove to capture hand pose and orientation in real time. Gesture commands are wirelessly transmitted via the nRF24L01 + module operating at 2.4 GHz, achieving a mean end-to-end latency of 8.3 +/- 1.4 ms. The robotic arm, actuated by four servo motors under a kinematic control scheme, achieves an average angular positioning accuracy of +/- 1.8 degrees, while the vehicle executes four directional commands with a motion accuracy of 95.3
Intelligent control systems are increasingly required for mobile robots operating under nonlinear wheel-terrain interaction, load redistribution, and uncertain terrain conditions. Although proportional-integral-derivative (PID) control is widely used because of its simplicity and practical implementation advantages, fixed-gain PID controllers are less effective when rover dynamics vary with slope, slip, and suspension articulation. This paper proposes a hybrid Deep Deterministic Policy Gradient-based adaptive PID control framework for a six-wheel rocker-bogie rover. The low-level PID loops maintain an interpretable control structure, while two decoupled DDPG agents tune the steering and drive PID gains online. A shaped reward function (DDPG-SR) is designed to penalize tracking errors, sudden gain changes, wheel mismatches, and physically inconsistent actuator responses. The controller is evaluated in a high-fidelity 3D MATLAB/Simscape Multibody environment and compared with fixed-gain PID and baseline DDPG-PID controllers. Simulation results show that the proposed DDPG-SR controller reduces the mean absolute steering error by approximately 81.0
Path planning and goal detection of a three Degrees of Freedom (DOF) moving robotic arm through a structured network environment with physical constraints is accomplished in this paper using a Reinforcement Learning (RL) approach. Considering the pattern of the network environments, the problem of path planning here is divided into a series of sub-goal detections. To achieve an intelligent control policy, the model-free Q-learning reinforcement algorithm and Deep Q-Networks (DQN) are employed and their efficiencies are analysed and compared. To show the efficiency of the proposed algorithms, a simulation environment inspired by the OpenAI Gym framework is designed and implemented, which accurately models the problem’s dynamics and its related physical constraints. Simulation results demonstrate the agent’s complete success in learning an optimal policy while showing superiority of DQN. Furthermore, qualitative evaluation of final trajectories shows that the agent is capable of finding optimal and efficient paths to perform assigned tasks while adhering to all constraints. Thus, it can be concluded that using the proposed reinforcement learning algorithms, an effective solution for complex robotic path detection problems can be successfully accomplished for industrial purposes involving scaffold environmental patterns.
Reactive Bug-based navigation algorithms remain widely used in mobile robotics as simple baseline strategies and as fallback behaviors within larger navigation systems, largely because they require minimal sensing and no global map. Despite their longevity, their limitations are typically described qualitatively or inferred indirectly from aggregate performance measures such as path length or success rate. A systematic account of how and why these algorithms fail in practice has remained largely absent. This study presents a quantitative failure-mode analysis of the canonical Bug algorithms Bug0, Bug1, and Bug2 under a unified experimental framework. The algorithms are evaluated over 135 runs, consisting of 15 start–goal cases across three deterministic, geometry-driven environments designed to expose interactions between algorithmic decision rules and obstacle structure. Rather than proposing a new navigation method, the paper introduces a reproducible taxonomy of failure behaviors based on log-detectable trajectory properties and execution events. The results show that failures are not random but arise through repeatable mechanisms linked to algorithm structure and environment geometry. Boundary over-commitment emerges as the dominant failure mode across all algorithms, indicating excessive adherence to obstacle boundaries. Bug0 primarily exhibits cyclic wandering characteristic of memoryless reactive navigation, while Bug1 frequently undergoes oscillatory re-entry, reflecting instability in boundary disengagement. Bug2 shows persistent boundary-dominated behavior, leading to inflated trajectories under constrained geometries. These behaviors are captured through measurable indicators including path inflation relative to a grid-based A* reference and boundary-domination ratios. The study provides a reproducible failure atlas that clarifies the operational limitations of Bug-based navigation and helps explain why such methods continue to be used as local or fallback components in modern robotic navigation architectures. The findings support more informed evaluation and deployment of reactive navigation strategies.
Multi-robot coverage path planning (mCPP) is critical for environmental monitoring, precision agriculture, and search-and-rescue operations. Existing methods often treat workload balancing and energy efficiency separately, leading to uneven task distribution and unnecessary energy consumption. This paper introduces a unified, energy-aware mCPP framework that integrates the Divide Areas Algorithm for Optimal Multi-Robot Coverage (DARP) with Turn-Minimized Spanning Tree Coverage (TMSTC). DARP provides equitable, connected region partitioning, while TMSTC generates low-turn trajectories that directly reduce energy use. We evaluate the framework through 336 experiments covering grid sizes from 20 × 20 to 50 × 50, obstacle densities of 10–15
This paper presents an event-triggered cascade active disturbance rejection control (CADRC) approach for quadrotor trajectory tracking. The proposed method decouples the system into position and velocity subsystems, with dedicated extended state observers (ESOs) and controllers designed for each subsystem to achieve accurate estimation and compensation of system nonlinearities, coupling effects, and external disturbances. To further optimize computational resource utilization, a hybrid event-triggering mechanism incorporating both static and dynamic thresholds is developed. This mechanism updates system states and control commands only at triggering instants, significantly reducing computational overhead. Additionally, an online self-tuning method based on particle swarm optimization (PSO) is introduced to enhance controller adaptability. Simulation results validate the robustness and efficiency of the proposed approach.
Autonomous mobile robot navigation in cluttered environments requires balancing path optimality with safety. Classical planners such as A* prioritize shortest paths but often produce trajectories with insufficient obstacle clearance, while approaches based on Artificial Potential Fields or scalarized costmaps suffer from local minima and loss of trade-off information. This paper introduces Costmap-MOA*, a multi-objective path planning framework that integrates continuous safety-aware costmaps with a Pareto-based search strategy. Unlike traditional methods, the proposed approach models safety as a continuous field and leverages an adapted NAMOA* algorithm to generate a dense set of non-dominated solutions. This enables explicit exploration of the distance–safety trade-off and reveals high-clearance paths that remain inaccessible to conventional planners. The framework is fully integrated into a ROS 2 architecture, ensuring practical applicability in real robotic systems. Simulation results obtained in both sparse and cluttered environments demonstrate that Costmap-MOA* significantly improves Pareto front density and quality, achieving up to a sixfold increase in non-dominated solutions and superior hypervolume and spread metrics compared to NAMOA* and NSGA-II. Additionally, comparisons with the ROS 2 Nav2 stack highlight its effectiveness in generating safer reference trajectories. These results confirm that the proposed method provides a robust and scalable solution for safety-aware autonomous navigation.
Wheeled jumping robots hold immense application potential in complex unstructured environments, such as industrial inspection and field exploration. However, vertical obstacles in complex terrains severely limit their traversability. As the core actuation unit for vertical obstacle negotiation, the configuration design of the jumping mechanism directly determines the robot’s stability during the energy storage phase and its maximum jumping height. Currently, the design of mainstream jumping mechanisms heavily relies on empirical trial-and-error paradigms, which often leads to poor energy storage stability and insufficient obstacle-clearing capabilities. To overcome this bottleneck, this paper proposes an optimization design framework tailored for jumping linkage mechanisms, based on the Spring-Connected Block Model (SBM). This framework not only achieves rapid topological optimization of the mechanism but also deeply integrates trajectory accuracy optimization with energy storage density optimization. Utilizing this method, a novel six-bar jumping mechanism featuring both high energy storage density and high-precision motion trajectories was successfully designed. Furthermore, the complete system design and performance validation of a wheeled jumping robot based on this mechanism were accomplished. Simulation results demonstrate that, compared to the Ascento four-bar robot of the same mass, the proposed novel mechanism increases the maximum jumping height by 21.5
This study proposes an energy-aware extension of the classical A* algorithm for quadrotor path planning in static and dynamic three-dimensional environments. Unlike distance-based A*, which mainly favors geometrically short paths, the proposed formulation incorporates quadrotor-relevant energy terms, including hover power, aerodynamic drag, climb energy, and propulsion-efficiency losses, directly into the path-evaluation process. The objective is to generate collision-free trajectories that reduce estimated flight energy while preserving the simplicity and interpretability of graph-search planning. The proposed planner was evaluated over 200 randomized simulation episodes, including 100 static-obstacle scenarios and 100 mixed static–dynamic scenarios. Its performance was compared with representative baseline planners, including distance-based A*, Theta*, Kinodynamic RRT*, and, in the static case, an energy-optimized three-dimensional planner. The evaluation considered energy consumption, path length, flight time, trajectory smoothness, planning time, and scalability. The results show that the proposed Energy-based A* planner reduced energy consumption compared with the evaluated baselines. Mean energy savings ranged from 21.40
Soft robotic grippers can handle delicate tasks but often lack tactile feedback for safe control. We present a 3D printed single-actuator three-finger Fin Ray inspired Flexi-Fin gripper with a laminated multilayer capacitive skin on each finger that provides real-time per-finger force sensing and slip onset detection from capacitance trends. The sensing layer is inexpensive and replaceable at about USD 0.20 per finger and fits into a shallow pocket that preserves compliance and avoids external housings. Our approach uses a simple on-board readout at high rate. Experiments show a linear force to capacitance calibration, gram-level resolution, 95
The issue with robotic path planning in dynamic environments also causes the problem to be a challenging issue, with nonlinear motion dynamics, uncertainty of the environment, and model inaccuracy being limiting. Under such circumstances, traditional control and model based approaches are frequently problematic, with regards to flexibility and real time decision making. Reinforcement Learning (RL) offers an alternative which is data-driven, which allows an agent to acquire navigation strategies during interaction with the environment. Soft Actor-Critic (SAC) is one of the RL approaches that perform well in continuous control problems due to its stability and exploration (based on entropy). Our proposed work discusses the developed modified SAC framework, which is also known as Memory-Constrained Adaptive Soft Actor-Critic (MCA-SAC) to improve learning efficiency and stability of path planning on robots. The algorithm incorporates adaptive priority of experience, dynamic reward shaping, and temporal action clustering into a memory constrained replay network. These adjustments enhance the use of samples, the speed of convergence, and the smoothness of policies when the obstacles are changed. The performance of the simulation shows that MCA-SAC converges quicker, nearly 10
Serial manipulators can have strongly varying internal stresses within their configuration space, making the identification of stress-critical poses important for structural design. Because each candidate configuration requires a computationally expensive Finite Element Analysis (FEA), high-fidelity simulations quickly become impractical. This paper proposes a Bayesian optimisation framework that searches directly over the robot configuration space, using peak Von Mises stress as the objective. For each candidate configuration, statically consistent interface reactions are computed and applied as boundary conditions in the structural model. A Gaussian process surrogate with an ARD Matérn 5/2 kernel models the configuration-stress relationship, while the Bayesian algorithm with Expected Improvement acquisition function selects new configurations for parallel evaluation. A case study on a cable-driven hyper-redundant robot demonstrates that the proposed workflow can identify configurations with critical stress states and accurately map the high-stress region with a limited number of simulations. The results show that the framework can support both structural verification and the detection of potentially critical robot poses during operation.
