Abstract The integration of large language models (LLMs) into KUKA industrial robots offers a promising pathway toward intuitive natural language control and a higher level of intelligence. However, existing methods face two primary challenges: their reliance on predefined motion primitives and the syntactic fragility imposed by the strict constraints of proprietary KUKA Robot Language (KRL). To address these challenges, we propose MIKI (multi-agent integrated KUKA interface), a decoupled dual-agent framework that bridges high-level reasoning and low-level syntactic verification for industrial robot programming. The controller agent incorporates a chain-of-thought (CoT) reasoning mechanism to systematically decompose abstract human instructions into logical intermediate steps. By leveraging a visual toolchain, it translates these reasoned plans into a sequence of executable actions, each embedded with precise world coordinates. Subsequently, the code generation agent employs a syntax-guided iterative refinement process, underpinned by an ANTLR4 parser, to transform these actions into verified and executable KRL code. Experimental results demonstrate that this closed-loop approach substantially improves the syntactic success rate of code generation from 0.61 to 0.89. Furthermore, the framework achieves an average success rate of 0.6 across five desktop manipulation tasks of varying complexity using a physical KUKA robotic arm, validating its effectiveness in bridging high-level cognitive reasoning with deterministic industrial execution.

Large position tracking errors and vibrations are commonly seen in robot machining due to the robot’s joint compliance. This paper proposes a dual-loop compensation control method to address this issue for robots with both motor and link side encoders. By taking advantage of link side encoders, the natural frequency and damping ratio of original robot dynamics are reshaped by a velocity compensation controller design to suppress vibrations. Meanwhile, link side position tracking error and velocity are employed to compute position compensation value to further improve robot machining accuracy. The effectiveness of this method was verified through single-joint simulation experiments. With large and noisy torques acting on the joint during robot machining, the simulated results showed that compared to full closed-loop P-PI control and ordinary damping control, the proposed method can significantly reduce the maximum position tracking error by up to 88.96

Under the demands of intelligent manufacturing and high-precision operations, robotic arm trajectory planning must ensure both motion smoothness and obstacle avoidance. Although search-based methods can quickly generate feasible paths, their trajectory smoothness is relatively low. In contrast, improved methods based on Bezier curves and optimization frameworks enhance smoothness, but often struggle with obstacle avoidance or computational inefficiency in complex environments. To address these issues, we propose a novel trajectory planning framework based on Velocity Adjustment Planning (VAP) in safe corridors. First, a collision-free corridor with geometric constraints is constructed by connecting key path points. Then, VAP introduces a dynamic modulation matrix to shape the velocity field in ellipsoidal regions, to enforce reachability constraints between arbitrary start and end. The global trajectory is constructed by progressively stitching together position and velocity segments. Simulation and real-world experiments are conducted on a 6-DOF RealMan robotic arm. The results demonstrate that the proposed method generates smooth and collision-free trajectories with strong robustness in complex environments.

Achieving human-like dexterity in robotic manipulation requires seamless integration of tactile perception and adaptive control strategies. This work presents a novel control framework integrating data-driven slip detection with constrained force optimization for four-fingered robotic manipulation using the Paxini Dexhand platform. We develop a comprehensive force balance constraint algorithm employing friction cone projections and optimization objective V(f) for real-time grasp force regulation. A compact slip detection network utilizing dual-domain feature extraction and channel attention mechanisms achieves high-precision slip identification, providing reliable tactile feedback for force control. The systematic integration strategy translates slip detection results into concrete force regulation actions, realizing a closed-loop system from reactive detection to proactive control. Experimental validation demonstrates detection latencies under 30ms with slip detection accuracy exceeding 95%, adaptively adjusting contact forces in real-time during dynamic manipulation tasks, reducing average gripping forces by 35-45% compared to conservative approaches while handling objects from delicate items to heavy industrial components.

Adaptive PID control plays an important role in trajectory tracking of robot manipulators. However, most theoretical researches on adaptive control exhibit certain discrepancies when applied to the practical robotic servo control system. This paper introduces a novel industrial-grade cascade adaptive Proportional-Proportional-Integral (PPI) controller based on a detailed analysis of the traditional PPI control architecture. Through rigorous Lyapunov stability analysis, the proposed adaptive PPI controller is proven to guarantee the asymptotic convergence of manipulator joint trajectory tracking errors to zero. Furthermore, it exhibits superior robustness against unknown loads and disturbances compared to classical PPI controllers and adaptive PD controllers. Extensive simulation studies and comparative analyses with scara robot manipulators validate the effectiveness and superiority of the proposed controller.

Dual-arm manipulation has gained significant attention in robotics, yet the safety of such systems remains an underexplored area of research. A key challenge is ensuring collisionfree operation while preserving effective coordination between the two arms. To address this challenge, we propose a novel framework for safe and efficient trajectory planning in dual-arm manipulation. The framework decomposes the dual-arm collaboration task into a low-dimensional quadratic programming (QP) problem. By solving this QP problem, the system ensures both physical constraints and collision avoidance, enabling successful dual-arm coordination. The proposed approach is evaluated through extensive simulations and physical experiments, demonstrating its robustness, safety, and effectiveness in real-world scenarios.

Quadratic programming methods have been widely used to solve the redundancy problem of control manipulators due to their ability to optimize performance indicators under physical constraints. However, conventional quadratic programming methods require known parameters. In practical work environments, the physical parameters of the manipulator can be uncertain. Therefore, an adaptive control method that can synchronize parameter identification and network control is necessary. We propose an adaptive varying parameter projection neural network method that begins from the perspective of joint velocity space. This method decouples the unknown parameters and nonlinear parts that require identification in the Jacobian matrix of the manipulator. For the first time, a projection network with varying parameters is introduced for redundant decomposition, greatly improving the accuracy of trajectory tracking. Additionally, we design an iterative identification equation based on neural dynamics and introduce position feedback, making Jacobian matrix identification more accurate. The proposed method can solve the problem of redundant resolution of robotic arms under actual physical constraints and unknown physical parameters and can improve performance indicators. Theoretical analysis and simulation results demonstrate the feasibility and performance of the proposed method.

Collision will result in mission failure or even damage to the robot. To address the mutual collision between redundant dual manipulators (RDMs), a novel double barrier function (DBF)-based mutual collision avoidance (MCA) scheme is proposed and investigated which can be formulated into a quadratic programming (QP)-based problem. Firstly, a novel safe barrier function (SBF)-based MCA inequality constraint is designed and derived which maximises the feasible region of collision avoidance and maintains the maximum safe distance between the RDMs under the same trajectory tracking task constraint. Secondly, a novel barrier varying-parameter recurrent neural network (BVRNN)-based QP solver with newly designed varying-parameter and activation function is proposed with faster convergence rate and higher error accuracy compared to the traditional varying-parameter neural network. Through the iteration and online-learning of the BVRNN-based QP solver, the RDMs can obtain the ability of MCA. Finally, Simulation experiments are presented to verify the effectiveness and superiority of the proposed DBF-based MCA scheme, i.e., ensuring that the RDMs are always at a set safe distance while performing a perfect end-effectors trajectory tracking task (error less than 10(-7)) during the collaborative operation process.

The stiffness and precision of robot joints are crucial for the processing performance of industrial robots. This study brings innovative works in aspects of joint structure design and control algorithms. Firstly, a dual-motor high-stiffness robot joint design with differential coupling reinforcement is proposed to increase transmission stiffness. Secondly, a frequency characterized disturbance rejection control method is introduced, in which disturbances with known frequency are observed and compensated accurately. Therefore disturbance’s influence on the system can be mitigated, which leads to enhancement of joint dynamic stiffness and tracking accuracy. Finally, the proposed joint design and control algorithm are validated through simulations. The results demonstrate that the proposed methods are able to significantly reduce speed ripples caused by disturbances with known frequency and improve trajectory tracking accuracy.

In this paper, a learning variable impedance control is designed for contact-rich robotics manipulation with guaranteed passivity. Firstly, a reinforcement-learning-based variable impedance policy is proposed, where the impedance parameters are selected as action space to achieve both the position precision and contact compliance. Since the RL-output impedance parameters may violate the passivity condition, a passivity-guaranteed impedance parameter generator is then designed to ensure the passivity of the robotics system. Additionally, an adaptive robust variable impedance control (ARVIC) framework is developed with the adaptive robust technique to cope with nonlinearity and uncertainties, which also ensures transiently convergence to the impedance contact model. Finally, a connector insertion task is designed to evaluate the effectiveness of the proposed system. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

Currently, in the field of industrial parts sorting, traditional methods for sorting metal parts often depend on manual labor, resulting in low efficiency and high costs. Conventional object detection algorithms exhibit high complexity However, in the production processes of many small and medium-sized enterprises, these algorithms often fail to meet the computational power required for object detection, posing challenges for implementing parts sorting solutions based on object detection. To address this challenge, we propose a novel object detection algorithm, YOLOv8n-VLAG, derived from YOLOv8n. The model incorporates the VanillaNet network as the backbone feature extraction network, significantly reducing the parameter count of YOLOv8n. Simultaneously, it introduces the LSKA attention mechanism, enhancing accuracy while reducing the parameter count. Additionally, GhostConv is introduced in the head section, further decreasing the parameter count while enhancing performance and maintaining nearly unchanged computational complexity. Extensive experiments were conducted on a self-curated industrial parts dataset. The model’s parameters and computational complexity amounted to $1.36 \times 10^{\wedge} 6$ and 4.5 GFLOPs, respectively, representing only $45.3 \%$ and $55.6 \%$ of yolov8n. The average accuracy experienced a slight improvement, while the mAP decreased by only $0.4 \%$, resulting in a reduction in the number of parameters and computational complexity at a minor expense. This study offers a viable solution for the detection of lightweight parts.

For most industrial/collaborative robot applications of model-based control, an accurate dynamic model is crucial to achieve good performance of the controller. Depending on the needs of different tasks, robots are often equipped with a variety of end effectors with various dynamic parameters (mass, center of mass and inertia), which could make the overall dynamics of the robot uncertain. This paper aims to identify the dynamic parameters of robot payload in its application by developing a new method with a 4-step motion, where only one joint needs to move in each step. Thanks to this particular motion with single joint, the robot dynamics can be decoupled and only the data of three joints which near the end-effector need to be collected. For each motion step, the adoption of a simplified dynamic model with fewer payload parameters is facilitated by the design of a special initial position and trajectory for a single joint, so that the impact of parameter on the accuracy of identification is significantly reduced compared with existing methods where multiple parameters are excited at the same time. Furthermore, a solving method of payload parameters based on the least squares method. The experimental results with a 6R industrial robot show the effectiveness of the proposed method for identifying different kinds of unknown payloads.

Structured light 3D imaging based on Gray code requires accurate pixel light and dark judgment for high-quality reconstruction. This paper proposes a pixel point classification method for fringe images based on the U-Net network to achieve robust and efficient pixel binary classification. To make the network more effective in learning and utilizing the structural information of the fringes projected by the projector, a multi-channel network input is designed. To address the challenge of obtaining a large number of fringe images for network training in real scenes, this paper utilizes computer graphics to construct a virtual 3D structured light imaging system and produce usable datasets. The proposed method achieved a pixel classification error rate of 0.44% and a mean square error of 285.019 on the test data, outperforming the traditional method's 1.10% and 713.578, respectively. The proposed method can accurately recognize and classify pixel values of highly reflective regions in an image. And it can generate high dynamic range images with wider dynamic range and richer image details.

Due to the combination of mobility and dexterity, the mobile manipulator has great application potential in many fields such as manufacturing, logistics and service. The motion planning for the mobile manipulator becomes more challenging due to the high degree of redundancy. An optimization-based method is proposed to solve the coordinated motion planning problem. The optimization goal is to enhance the manipulability while considering the distinct kinetic characteristics of the manipulator and the mobile platform. Since the manipulator has superior performance in respond speed and accuracy, a motion distribution weighting matrix is employed to increase the motion proportion of the manipulator for a given task. Moreover, the joint range and velocity constraints are formulated in a unified form for calculation efficiency. Simulation and experiment are presented to validate the proposed method. The algorithm can generate feasible trajectories in configuration space without violating the constraints, and the adjustable motion distribution between the manipulator and the mobile platform is achieved. The trajectory tracking accuracy can be improved when the configuration space motion distribution between the manipulator and the mobile platform is regulated appropriately.

Physical feasibility constraints play an important role in robot dynamics parameter identification. However, in practical robot development, not only physical feasibility is required, but also mapping the real inertial properties of each link. In this work, the latter requirement is called physical reality constraints. To address this problem, a two-step identification method for identifying the complete set of inertial parameters is adopted to guarantee the identified result is optimal in both static and dynamic environments while considering physical reality. To fulfill physical reality constraints, the dynamic parameters retrieved from the robot CAD model are used as the initial guesses in the optimization process, and the parameters’ lower and upper boundaries are decided by adding and subtracting a suitable value respectively. The proposed approach is validated on a six-DOF collaborative robot.

Mobile manipulator is increasingly expected to undertake tasks in intelligent automated factories. However, when facing resistance tasks in small spaces, it is a challenge to ensure the safety of target objects and robots without using high-precision positioning sensors. To address this issue, we propose a force exertion method for redundant mobile manipulators. To improve force exertion capability, the force manipulability ellipsoid of the mobile manipulator is optimized by exploiting its redundancy under a whole-body controller. Furthermore, a hybrid force-impedance control strategy is developed to guarantee safe operation in compact-space resistance tasks. The proposed method is validated by performing a turn-on-switch experiment.

为了获取柔性关节精确的物理参数,提出了一种基于系统谐振与抗谐振特性的参数辨识方法.首先建立柔性关节的数学模型,利用该模型推导柔性关节的谐振、抗谐振频率特性与待辨识参数的数学关系,然后基于此关系建立误差回归模型,设计实验采集不同负载条件下的输入与输出数据,计算得到系统的谐振、抗谐振频率及幅值,代入回归模型并基于最小二乘法求解参数.最后,通过仿真与实验将本文方法与一般的频域特性拟合方法进行对比,结果表明在含有噪声的情况下本文方法将参数辨识平均精度从75.34％提高到90.35％,方差从25.34％降低到8.07％,验证了所提方法的可行性与有效性.

重力补偿方法广泛地应用于由连杆与旋转机构组成的机器人系统中,更换机器人末端执行器造成了补偿模型的不确定性.针对该问题,提出了一种利用机器人关节力矩与位置信息的负载参数离线辨识方法.基于机器人静力学方法提出了2种负载参数的计算模型,并通过采集机器人在多个静态位姿条件下的关节力矩与位置信息获得负载参数的最小二乘解.进一步,本文针对机器人的辨识位姿选取问题展开研究,提出了同时保证辨识精度与辨识简便性的多目标优化问题,使用多目标粒子群优化方法获得最优辨识位姿.根据辨识后的负载参数,给出了机械臂各关节负载的重力补偿量计算方法.实验结果表明所提方法具有较高的辨识精度,负载质量的辨识误差最小值达到0.00706 kg,最大值达到0.151 kg,负载质心位置的辨识误差最小值达到0.0254 m,最大值达到0.122 m,验证了上述方法的可行性与有效性.

In this paper, an impedance control method based on backstepping approach is proposed aiming at the problems of compliance control when the flexible joint contacts with the environment. The stability of the controller is proved based on Lyapunov stability theory. The controller design combines stability of a Lyapunov function and the desired dynamic of impedance model. The experiment results show the effectiveness and feasibility of the proposed control approach.