Multi-agent reinforcement learning (MARL) remains fundamentally challenged by partial observability, unstable value learning, and inefficient exploration—difficulties that intensify in high-dimensional robotic control and large-scale coordination scenarios. Meanwhile, the existing algorithms lack a mechanism to guide the improvement of long-term strategies. We propose PRIME, Policy Representation Integration with Meta-Value–Modulated Evolution in Multi-Agent Reinforcement Learning that addresses these limitations through representation-asymmetric policy parameterization, meta-value–augmented optimization, and meta-value–modulated evolutionary search. PRIME constructs a shared nonlinear encoder with lightweight team-specific linear heads, providing a coherent latent policy manifold that supports both fine-grained robotic manipulation and large-population coordination. A learned meta-value function estimates the long-horizon utility of policy updates, whose gradients shape both actor learning and representation formation. In parallel, evolutionary operators—direction-aware crossover and meta-gradient–scaled low-rank mutation—enable globally diverse yet strategically targeted exploration in policy space. Evaluations on Multi-Agent MuJoCo, DexHands dexterous manipulation, and the large-scale DCA benchmark demonstrate that PRIME achieves consistently superior performance, faster convergence, and stronger robustness than state-of-the-art base-lines.

The Linear Quadratic Regulator (LQR) problem for systems with uncertainty is challenging: model-based design loses optimality, while prevailing reinforcement learning methods demand prohibitive data and computation. This paper introduces a Hybrid-Driven Recurrent Learning (HDRL) framework that bridges this gap by synergizing model-based optimal control with data-driven learning in a novel way. The core of HDRL is a hybrid training strategy: the policy is evaluated by rolling out on the actual uncertain system (forward pass), while the policy gradient is computed by backpropagating through a deterministic nominal model (backward pass). This approach creates a low-variance, model-guided policy gradient, a fundamental departure from high-variance model-free estimators. Architecturally, HDRL employs a recurrent neural network-like structure, repurposing the LQR cost as a self-supervised loss. A key enabler is our method to convert both additive and structured uncertainties into an additive signal, making training feasible without knowledge of the perturbed dynamics. Simulations demonstrate that HDRL provides a robust, sample-efficient, and practical solution for optimal control under uncertainty.

Due to the density inconsistency and distribution difference between cross-source point clouds, previous methods fail in cross-source point cloud registration. We propose a density-robust feature extraction and matching scheme to achieve robust and accurate cross-source registration. To address the density inconsistency between cross-source data, we introduce a density-robust encoder for extracting density-robust features. To tackle the issue of challenging feature matching and few correct correspondences, we adopt a loose-to-strict matching pipeline with a ``loose generation, strict selection'' idea. Under it, we employ a one-to-many strategy to loosely generate initial correspondences. Subsequently, high-quality correspondences are strictly selected to achieve robust registration through sparse matching and dense matching. On the challenging Kinect-LiDAR scene in the cross-source 3DCSR dataset, our method improves feature matching recall by 63.5 percentage points (pp) and registration recall by 57.6 pp. It also achieves the best performance on 3DMatch, while maintaining robustness under diverse downsampling densities.

Object detection in aerial images has always been a challenging task due to the generally small size of the objects. Most current detectors prioritize the development of new detection frameworks, often overlooking research on fundamental components such as feature pyramid networks. In this paper, we introduce the Cross-Layer Feature Pyramid Transformer (CFPT), a novel upsampler-free feature pyramid network designed specifically for small object detection in aerial images. CFPT incorporates two meticulously designed attention blocks with linear computational complexity: Cross-Layer Channel-Wise Attention (CCA) and Cross-Layer Spatial-Wise Attention (CSA). CCA achieves cross-layer interaction by dividing channel-wise token groups to perceive cross-layer global information along the spatial dimension, while CSA enables cross-layer interaction by dividing spatial-wise token groups to perceive cross-layer global information along the channel dimension. By integrating these modules, CFPT enables efficient cross-layer interaction in a single step, thereby avoiding the semantic gap and information loss associated with element-wise summation and layer-by-layer transmission. In addition, CFPT incorporates global contextual information, which improves detection performance for small objects. To further enhance location awareness during cross-layer interaction, we propose the Cross-Layer Consistent Relative Positional Encoding (CCPE) based on inter-layer mutual receptive fields. We evaluate the effectiveness of CFPT on three challenging object detection datasets in aerial images: VisDrone2019-DET, TinyPerson, and xView. Extensive experiments demonstrate that CFPT outperforms state-of-the-art feature pyramid networks while incurring lower computational costs. The code is available at https://github.com/duzw9311/CFPT.

As a crucial component of object detectors, current detection heads often lack the capability to effectively utilize contextual information, adapt to deformable objects, and align features and tasks. However, most existing methods prioritize a single capability, lacking comprehensive approaches to introduce them simultaneously. In this paper, we propose the Enhanced Head to integrate the above three capabilities into the detectors concurrently. Specifically, we propose three attention blocks with linear complexity: Global Concentrated Attention (GCA), Local Deformable Cross-Task Attention (LDCA), and Boundary-Aware Cross-Task Attention (BACA). The GCA captures long-range dependencies efficiently by employing Spatial Information Concentration (SIC). The LDCA improves feature alignment and deformation adaptability by enabling local deformable cross-task feature interactions. The BACA aligns classification features with localization results, enhancing task alignment and further improving deformation adaptability through a region-deformable interaction scheme. We implement Enhanced Head as a plug-and-play detection head and evaluate its effectiveness through extensive experiments on the MS COCO and VisDrone datasets. For instance, on the COCO detection benchmark, our Enhanced Head achieves +3.6 AP gain for FSAF, +3.3 AP for RetinaNet, and +2.9 AP for ATSS while reducing the FLOPs.

In this paper, a hypervolume contribution approximation method is proposed. The main idea is to find out all angular points in the hypervolume contribution region of the solution by iteration and make an approximation by the distance of the line segment from angular points to reference points. The searching strategy is introduced first, and then the angular points elimination/compensation measures are proposed. In the experiment section, we introduce two state-of-the-art hypervolume contribution approximation methods for comparison with the proposed algorithm. These methods are tested on six different Pareto front solution sets. The results show that the proposed method has better performance than other methods. In particular, the proposed method is competitive in recognizing the hypervolume contribution of large-scale solution sets.

Addressing the challenges posed by the substantial gap in point cloud data collected from diverse sensors, achieving robust cross-source point cloud registration becomes a formidable task. In response, we present a novel framework for point cloud registration with broad applicability, suitable for both homologous and cross-source registration scenarios. To tackle the issues arising from different densities and distributions in cross-source point cloud data, we introduce a feature representation based on spherical voxels. Furthermore, addressing the challenge of numerous outliers and mismatches in cross-source registration, we propose a hierarchical correspondence filtering approach. This method progressively filters out mismatches, yielding a set of high-quality correspondences. Our method exhibits versatile applicability and excels in both traditional homologous registration and challenging cross-source registration scenarios. Specifically, in homologous registration using the 3DMatch dataset, we achieve the highest registration recall of 95.1% and an inlier ratio of 87.8%. In cross-source point cloud registration, our method attains the best RR on the 3DCSR dataset, demonstrating a 9.3 percentage points improvement. The code is available at https://github.com/GuiyuZhao/VRHCF.

In this paper, we introduce a new outlier removal method that fully leverages geometric and semantic information, to achieve robust registration. Current semantic-based registration methods only use semantics for point-to-point or instance semantic correspondence generation, which has two problems. First, these methods are highly dependent on the correctness of semantics. They perform poorly in scenarios with incorrect semantics and sparse semantics. Second, the use of semantics is limited only to the correspondence generation, resulting in bad performance in the weak geometry scene. To solve these problems, on the one hand, we propose secondary ground segmentation and loose semantic consistency based on regional voting. It improves the robustness to semantic correctness by reducing the dependence on single-point semantics. On the other hand, we propose semantic-geometric consistency for outlier removal, which makes full use of semantic information and significantly improves the quality of correspondences. In addition, a two-stage hypothesis verification is proposed, which solves the problem of incorrect transformation selection in the weak geometry scene. In the outdoor dataset, our method demonstrates superior performance, boosting a 22.5 percentage points improvement in registration recall and achieving better robustness under various conditions. Our code is available.

Feature upsampling is an essential operation in constructing deep convolutional neural networks. However, existing upsamplers either lack specific feature guidance or necessitate the utilization of high-resolution feature maps, resulting in a loss of performance and flexibility. In this paper, we find that the local self-attention naturally has the feature guidance capability, and its computational paradigm aligns closely with the essence of feature upsampling (i.e. feature reassembly of neighboring points). Therefore, we introduce local self-attention into the upsampling task and demonstrate that the majority of existing upsamplers can be regarded as special cases of upsamplers based on local self-attention. Considering the potential semantic gap between upsampled points and their neighboring points, we further introduce the deformation mechanism into the upsampler based on local self-attention, thereby proposing LDA-AQU. As a novel dynamic kernel-based upsampler, LDA-AQU utilizes the feature of queries to guide the model in adaptively adjusting the position and aggregation weight of neighboring points, thereby meeting the upsampling requirements across various complex scenarios. In addition, LDA-AQU is lightweight and can be easily integrated into various model architectures. We evaluate the effectiveness of LDA-AQU across four dense prediction tasks: object detection, instance segmentation, panoptic segmentation, and semantic segmentation. LDA-AQU consistently outperforms previous state-of-the-art upsamplers, achieving performance enhancements of 1.7 AP, 1.5 AP, 2.0 PQ, and 2.5 mIoU compared to the baseline models in the aforementioned four tasks, respectively.

Abstract Most multiple object tracking algorithms depend on the output of the detector. Aiming at the problem that the higher detection quality model is restricted by the computing power, and the robustness of the lightweight detection model is easily affected by motion blur, this paper proposes a lightweight moving object detector based on improved YOLOv8 combined with dynamic confidence compensation algorithm. The algorithm combines various technical means such as network structure optimization, lightweight design, self‐knowledge distillation, loss function improvement and dynamic confidence compensation. ByteTrack is used as a tracker to conduct experiments on PASCAL VOC07+12 data set and UA‐DETRAC test sequence. Compared with the baseline YOLOv8n+ByteTrack, the proposed algorithm improves the HOTA by 1.3% when the single frame tracking delay is reduced by 1.1%. Mostly tracked target is improved by 79.7%, mostly lost target is reduced by 10.9%, and the detection effect is better than the original detector and other popular object detectors. The YOLODCC model achieves a balance between lightweight and multi‐object motion blur.

Point cloud registration aims to estimate a transformation that aligns point clouds collected from different perspectives. In learning-based point cloud registration, a robust descriptor is crucial for achieving high-accuracy registration. However, most existing methods are susceptible to noise and demonstrate poor generalization ability when applied to unseen datasets. Motivated by this, we introduce SphereNet to learn a noise-robust and unseen-general descriptor for point cloud registration. In our method, first, the spheroid generator builds a geometric domain based on spherical voxelization (SV) to encode geometric information. Then, the spherical interpolation of the sphere is introduced to realize robustness against noise. Finally, a new spherical convolutional neural network (CNN) with spherical integrity padding completes the extraction of descriptors, which reduces the loss of features and fully captures the geometric features. To evaluate our methods, a new benchmark 3DMatch-noise with strong noise is introduced. Extensive experiments are carried out on both indoor and outdoor datasets. Our results demonstrate that SphereNet achieves an increase in feature-matching recall of more than 25 percentage points (pp) on 3DMatch-noise under high-intensity noise. Moreover, SphereNet establishes a new state-of-the-art performance on the 3DMatch and 3DLoMatch benchmarks, achieving 93.5% and 75.6% registration recall (RR), respectively. Furthermore, SphereNet exhibits superior generalization ability on unseen datasets.

Obstacle detection plays an important role for robot collision avoidance and motion planning. This paper focuses on the study of the collision prediction of a dual-arm robot based on a 3D point cloud. Firstly, a self-identification method is presented based on the over-segmentation approach and the forward kinematic model of the robot. Secondly, a simplified 3D model of the robot is generated using the segmented point cloud. Finally, a collision prediction algorithm is proposed to estimate the collision parameters in real-time. Experimental studies using the Kinect Ⓡ sensor and the Baxter Ⓡ robot have been performed to demonstrate the performance of the proposed algorithms.

In this paper, a neural networks (NNs) enhanced telerobot control system is designed and tested on a Baxter robot. Guaranteed performance of the telerobot control system is achieved at both kinematic and dynamic levels. At kinematic level, automatic collision avoidance is achieved by the control design at the kinematic level exploiting the joint space redundancy, thus the human operator would be able to only concentrate on motion of robot's end-effector without concern on possible collision. A posture restoration scheme is also integrated based on a simulated parallel system to enable the manipulator restore back to the natural posture in the absence of obstacles. At dynamic level, adaptive control using radial basis function NNs is developed to compensate for the effect caused by the internal and external uncertainties, e.g., unknown payload. Both the steady state and the transient performance are guaranteed to satisfy a prescribed performance requirement. Comparative experiments have been performed to test the effectiveness and to demonstrate the guaranteed performance of the proposed methods.

A lot of real-life data are of graph nature. However, it is not until recently that business begins to exploit data's connectedness for business insights. On the other hand, RDBMSs are a mature technology for data management, but they are not for graph processing. Take graph traversal, a common graph operation for example, it heavily relies on a graph primitive that accesses a given node's neighborhood. We need to join tables following foreign keys to access the nodes in the neighborhood if an RDBMS is used to manage graph data. Graph exploration is a fundamental building block of many graph algorithms. But this simple operation is costly due to a large volume of I/O caused by the massive amount of table joins. In this paper, we present G-SQL, our effort toward the integration of a RDBMS and a native in-memory graph processing engine. G-SQL leverages the fast graph exploration capability provided by the graph engine to answer multi-way join queries. Meanwhile, it uses RDBMSs to provide mature data management functionalities, such as reliable data storage and additional data access methods. Specifically, G-SQL is a SQL dialect augmented with graph exploration functionalities and it dispatches query tasks to the in-memory graph engine and its underlying RDMBS. The G-SQL runtime coordinates the two query processors via a unified cost model to ensure the entire query is processed efficiently. Experimental results show that our approach greatly expands capabilities of RDBMs and delivers exceptional performance for SQL-graph hybrid queries.

In this paper, a human robot shared control strategy is developed and tested on a Baxter robot. Using the proposed method, the human operator only needs to consider the motion of the end-effector of the manipulator, while the manipulator will avoid obstacle by itself without sacrificing the end effector motion performance. An improved obstacle avoidance strategy based on the joint space redundancy of the manipulator is designed. A dimension reduction method is presented to solve the over defined problem of avoiding velocity to achieve a more efficient use of the redundancy. By employment of an artificial parallel system of the teleoperate manipulator and the task switching weighting factor, the proposed control method enable the robot restoring back to the commanded pose smoothly when the obstacle is removed. By implementing the dimension reduction method, the trajectory of each joint of the manipulator can be controlled at the same time to achieve the restoring task. Thus, the proposed control method can eliminate the impact of the obstacle on the remaining task. Satisfactory experiment results demonstrate the effectiveness of the proposed methods.

By the cooperative and competitive PSO algorithm, the goal of this study is to provide the cooperative trajectories of multiple UAVs in the three dimensional space. To effectively reduce the dimension of this problem, the optimization process is mainly divided into two substages to reduce the difficulty of selecting the weights of objectives and constraints in the considered objective function. Considering several objectives and constraints, the cooperative trajectories in the first substage are given by the cooperative and competitive PSO algorithm in the two dimensional space. In the second substage, the altitude of cooperative trajectories is adjusted according to the considered objectives and constraints. In the complicated scenarios, simulation results demonstrate the effectiveness and the robustness of the cooperative and competitive PSO algorithm, which possibly provides one guideline for optimal cooperative planning trajectories of multiple UAVs in the three-dimensional space.

This paper describes our preliminary development of a virtual robot teleoperation platform based on hand gesture recognition using visual information. Hand gestures in images captured by a camera are recognised to control a virtual iCub. We employ two methods to realise the classification: Adaptive Neuro-fuzzy Inference Systems (ANFIS) and Support Vector Machines (SVM). We realise the teleoperation of a virtual robot using iCubSimulator. The technique in the paper will enable us to teleoperate a physical robot in the future work. In addition, a video server is set up to monitor the real robot. By using the parallel system we are able to improve the robot's performance. Based on the techniques presented in this paper, the virtual iCub can perform the specified actions remotely in a natural manner.

In this paper, in order to combine intelligence of human operator and automatic function of the robot, we design a control scheme for the bimanual robot manipulation, in which the leading robot arm is directly manipulated by a human operator through a haptic device and the following robot arm will automatically adjust its motion to match the operator's motion. In this paper, we propose a fuzzy-based adaptive feedforward compensation controller and apply it into the robot control. According to the comparison results in the simulated experiment, we conclude that the fuzzy-adaptive controller performs better than the non-fuzzy controller, although they can both complete the specified task by tracking the leading robot arm controlled by the human operator. The techniques developed in this paper could be very useful for our future study on adaptation in human-robot interaction in improving the reliability, safety and intelligence.

Recently, Honda Eco Mileage Challenge, which is one of the most important contests in the world to encourage high-efficient car and environmental protection, mainly has been attracted by the communities of mechanism engineering, control system, vehicle engineering and computer science, etc. Generally speaking, the final score of HEMC contest is mainly related to the number of the firing in the whole process. In essence, the firing planning problem is a typical constraint optimization problem with several objectives. In order to obtain the good score in the HEMC contest, the PSO algorithm is utilized to optimize the number of the firing, the concrete firing time and the concrete position of each track in the whole process. To demonstrate the effectiveness and high performance of PSO algorithm, numerical results in the firing planning problem can help the driver to provide the firing time and the concrete firing position in the F1 racing track.

Formation control problem has been extensively investigated in the literature of multi-agent systems, robotics, and control, etc. Our previous work mainly concentrated on the theoretic study of line formation with three robots, however, it is hard to handle with the formation problem whose number of robots is strictly larger than 3. In order to effectively overcome this problem, this paper incorporates convex hull and the standard PSO algorithm to design the typical formation of several robots. Firstly, on the basis of convex hull of robots, objective function, corresponding to several constraints, is given by the new convex hull method. Secondly, the standard PSO algorithm is adopted to search for the desired positions of several robots to minimize the objective function and satisfy the formation constraints. To demonstrate the effectiveness of the proposed algorithm, numerical results, regarding the formation of several ships in the realistic ocean, mainly concentrate on triangle formation, diamond formation and regular polygon formation.

Wireless sensor network localization is an essential problem that has attracted increasing attention due to wide requirements such as in-door navigation, autonomous vehicle, intrusion detection, and so on. With the a priori knowledge of the positions of sensor nodes and their measurements to targets in the wireless sensor networks (WSNs), i.e. posterior knowledge, such as distance and angle measurements, it is possible to estimate the position of targets through different algorithms. In this contribution, two approaches based on least-squares and Kalman filter are described for localization of one static target in the WSNs with distance, angle, or both distance and angle measurements, respectively. Noting that the measurements of these sensors are generally noisy of certain degree, it is crucial and interesting to analyze how the accuracy of localization is affected by the sensor errors and the sensor network, which may help to provide guideline on choosing the specification of sensors and designing the sensor network. To this end, we make theoretical analysis for the different methods based on three types of measurement noise: bounded noise, uniformly distributed noises, and Gaussian white noises. Simulation results illustrate the performance comparison of these different methods.

3D entity model with the data structure of regular hexahedron, which is usually called as 3D regular block model, is now widely used in many domains such as resources estimation (RE) and finite element analysis (FEA) etc. In these domains, each regular hexahedron element of a 3D entity should be respectively evaluated with spatial related properties. However, 3D entity models are usually constructed from limit and sparse geometrical elements such as feature points and contour line strings, so these models are usually represented only with connected triangles as their surface but nothing in their hollow interior. We need a method to tessellate these models into regular blocks and fill the hollow interior with them. This paper mainly introduces an effective tessellation algorithm for converting 3D wireframe models which are represented as a collection of surface triangles to 3D regular block models which are represented as a collection of regular hexahedrons. With this algorithm, both simple 3D entity model with single outside boundary and complex 3D entity model with inside cavities can be tessellated into a collection of regular hexahedrons, which are constrained to the inside and outside boundaries of 3D entities. We have developed a test application for resources estimation based on this algorithm. A gold ore-body wire-frame model is tessellated into regular hexahedrons. The inverse distance weighting (IDW) interpolation method is used to evaluate each regular hexahedron with gold grade. The 3D block model of this gold ore-body is visualized with Au-grade distribution information.