To address the streaming arrival of point cloud data and the widespread sparsity and incompleteness of point clouds in real-world scenarios, this paper proposes a dual-stream interaction and generative replay framework for lifelong point cloud understanding, termed RADR-Point. The framework aims to tackle two core challenges: insufficient feature representation caused by sparse point clouds and catastrophic forgetting during incremental learning. Overall, we design a multi-task collaborative network to achieve deep coupling between point cloud classification and completion. The model employs a semantic gated fusion mechanism to adaptively inject high-level category attributes from the classification stream into the completion stream, thereby guiding the plausible reconstruction of locally missing structures with accurate semantic priors. Meanwhile, high-quality geometric feedback also significantly enhances the representational robustness of classification features. To address the inevitable forgetting problem in sequential tasks, we construct a reward-aware dynamic rethinking mechanism based on a conditional latent diffusion model. This mechanism selects historical feature anchors with high information entropy by jointly evaluating cognitive uncertainty and reconstruction difficulty, and leverages cross-attention to precisely guide the diffusion denoising process. As a result, it effectively mitigates mode collapse and recovers historical feature manifolds with high diversity and fidelity for participation in dual-stream interaction. Experiments on continual task benchmarks show that the proposed framework improves classification accuracy and completion quality compared with the evaluated baselines.

Unmanned aerial–aquatic vehicles (UAAVs) provide cross-domain adaptability and broad visions, while autonomous underwater vehicles (AUVs) support long-duration operations. This work integrates the two by developing a rapid underwater docking and releasing system. An autonomous clamping mechanism is designed to anchor UAAVs under varied landing attitudes, and a vision–tactile state perception algorithm based on decision-level dual-modal fusion is proposed to enable reliable underwater docking with no need of communications between the UAAV and AUV. Experimental results validate autonomous perception and reliable docking in fully underwater environments, achieving a docking time of 6 s and a landing gear recognition accuracy of 3 mm. The proposed framework offers an efficient solution for aerial–aquatic cooperation, advancing cross-domain robotic platforms for ocean monitoring, emergency response, and underwater exploration.

The large-scale deployment of unmanned aerial vehicles (UAVs) in low-altitude airspace requires efficient and reliable conflict resolution methods. This paper presents a comparative study of mixed-integer linear programming (MILP) and mixed-integer nonlinear programming (MINLP) formulations for heading-only UAV conflict resolution under three representative spatial configurations: circular, diamond, and random. To ensure a fair comparison, both formulations are implemented using the SCIP optimization framework. Extensive simulations with varying fleet sizes demonstrate that computational performance is governed primarily by spatial configuration rather than optimization class. In the structured diamond configuration, both models achieve comparable solution quality, with MINLP providing a modest reduction in computation time. In the symmetric circular configuration, MINLP consistently outperforms MILP by avoiding discretization-induced inefficiencies. In the unstructured random configuration, MINLP exhibits substantial advantages in both runtime and solution quality, whereas MILP experiences rapidly increasing computational complexity. These findings indicate that nonlinear optimization is not inherently less efficient than linear optimization for UAV conflict resolution. Instead, the underlying spatial conflict structure plays a decisive role in algorithm performance, providing practical guidance for selecting optimization frameworks in future UAV traffic management systems.

Leveraging large language models (LLMs) for collaborative sequential decision-making is a significant challenge, despite strong semantic understanding and extensive prior knowledge. Conversely, multi-agent reinforcement learning (MARL) can learn environment-aligned policies through interaction, but often suffers from inefficient exploration and heavy reliance on centralized global state information. To achieve complementary advantages, we propose the environment-aligned LLMs (EALLMs). In our framework, an LLM serves as a shared policy for all agents and is updated through online MARL to achieve alignment with the environment. Simultaneously, another LLM, fine-tuned with offline datasets, acts as an information integrator to generate global state for communication purposes. Additionally, we design robust, task-specific prompts tailored to multi-agent systems. Extensive experiments demonstrate that EALLMs outperform classical MARL and LLM-based baselines in both exploration efficiency and overall performance on the SMAC and SMACv2 benchmarks. Ablation studies further confirm EALLMs’ ability to achieve competitive results without relying on explicit global state, while preserving the original capabilities of the LLM during alignment.

Complex multi-agent control tasks remain challenging for traditional rule-based and model-based approaches, motivating the adoption of learning-based methods. However, learning-based methods often struggle with sim-to-real transfer because they rely on accurate dynamics modeling or system identification and learn policies in low-level control spaces that are highly sensitive to dynamics mismatch, making them costly and fragile in complex environments. To address this issue, we propose a sim-to-real method for multi-agent control, which is insensitive to dynamics mismatch via effect alignment. Our method combines random environmental structure with discrete semantic actions through closed-loop control, elevating policy learning to a semantic abstraction level. Additionally, we develop an action synchronization mechanism that mitigates inter-agent action timing mismatches, thereby enhancing the temporal consistency of the system. Experiments on four multi-agent navigation tasks demonstrate that our method substantially improves training efficiency over mainstream transfer methods and achieves higher success rates in real-world scenarios, thereby improving the robustness and deployment stability of multi-agent systems under dynamics mismatch.

Emerging mobile IoT, augmented/virtual reality (AR/VR) and robotic applications require continuous wireless power delivery together with accurate spatial awareness. Millimeter-wave (mmWave) array beamforming can enhance power focusing and angular resolution, but it usually requires stringent beam control, phase alignment, and feedback signaling. To address these challenges, this paper proposes a resonant beam-based power and direction-sensing system (RB-PDS), which employs retrodirective antenna arrays (RAAs) and closed-loop resonant propagation to establish a self-aligned focused beam between the base station (BS) and an RF-passive user device (UD). The same resonant beam is reused as both the energy-transfer carrier and the direction-sensitive sensing signal, thereby enabling simultaneous wireless power transfer (WPT) and direction-of-arrival (DoA) estimation without independent localization waveforms or active RF transmission from the UD. To improve DoA accuracy in the Fresnel range, a two-step modified multiple signal classification (M-MUSIC) algorithm is developed to compensate for the wavefront curvature of the converged beam. Analytical modeling and system-level simulations show that under the considered short-range operating conditions, RB-PDS can deliver watt-level net DC power while achieving centimeter-level lateral localization accuracy. These results demonstrate the potential of RB-PDS as a low-overhead framework for integrated wireless powering and direction sensing in short-range intelligent systems.

Vision based and event based tactile sensors are important in robotic manipulation research. However, they suffer from a fundamental tradeoff: vision based sensors have low sampling rates, while event based sensors are prone to drift during long term static force estimation. To solve this challenge and achieve human level tactile perception, the novel hybrid event frame tactile sensor (Mixtac) is proposed in this paper by emulating the synergistic function of biological mechanoreceptors, which achieves normal force estimation. The prototype leverages events for high frequency force tracking and frames for long term accuracy. The Frame Guided Event Recurrent Network (FGER-Net) was proposed to fuse the two data streams. Frames were used by the net to correct event drift during training and guide high frequency predictions during inference. Experiments demonstrated an MAE of 0.04 N. This paper could bridge the sampling rate gap from 0 to 500 Hz in current vision based tactile sensors and pave the way for human level robotic manipulation.

Cooperative tasks are common in multi-agent systems, with closely cooperative tasks being a special case of this, where a change in the state of the environment requires multiple agents to perform a specific operation at the same time. Take a box-pushing task as an example, the box is heavy and requires multiple agents to push it simultaneously. Optimal actions in a closely cooperation task are correlated with the actions of other agents, which makes the individual optimal action potentially inconsistent with the group optimal action, which leads to more non-globally optimal Nash equilibrium policies in the problem. This makes it easier for the policy learned by reinforcement learning to fall into these locally optimal policies. In this paper, we propose a self-organised sequential multi-agent reinforcement learning algorithm (SOS-MARL). We propose sequential decision-making to change the optimization objective of the agent's policy so that the learned policy tends to group optimal policies. And propose an automatic grouping mechanism to make the policy smoother for training and reasoning in large-scale agent environments. We decompose the joint action value factorization outside the group into a combination of each group action value, thus guiding the agents to improve their group policies in a fine-grained manner. We deployed scenarios in both simulated and real environments and compared SOS-MARL with various classical MARL algorithms on box-pushing tasks, demonstrating the state-of-the-art of our method.

Efficiently and reliably determining all natural frequencies below a trial frequency is critical for analyzing the vibration characteristics of mechanical systems. The multibody system transfer matrix method (MSTMM) is a powerful tool for mechanical system dynamics analysis. The derived eigen-equation contains nonlinear functions of the eigenvalues, making the conventional methods (such as subspace iteration, QR iteration, etc.) for solving the generalized eigenproblem not applicable. The low order of its coefficient matrix makes it feasible to solve the equivalent nonlinear algebraic equation to determine the natural frequencies, but many methods for solving nonlinear equations locally converge, making it difficult to obtain all natural frequencies below a trial frequency. To overcome this difficulty, based on MSTMM, this paper proposes an improved automatic search method and its acceleration to efficiently determine all natural frequencies below a trial frequency. Firstly, the iterative step error formula of the square root iteration method which is globally and cubically convergent is modified to avoid the computational failure due to too small or large values and tedious calculation caused by the coefficient matrix determinant and its derivatives, thereby enhancing the computational efficiency and stability. Secondly, an iterative initial value estimation method is proposed to obtain an initial value iterating to the next natural frequency after already obtaining one natural frequency, where the required natural frequency multiplicity can be obtained by the proposed multiplicity determination method. Thirdly, a detection method is proposed to avoid missing natural frequencies due to root over-stepping. These contributions lead to an improved automatic search method (IASM). Further, the numerical deflation is introduced to accelerate the IASM, resulting in an improved accelerated automatic search method. Finally, the efficiency and reliability of the proposed methods are verified through comparative numerical simulations. The proposed methods are highly efficient, automatically performed and reliable.

The integration of virtual and real environments is driving transformative advancements in unmanned systems, offering new avenues for intelligent collaboration and adaptive deployment. This survey presents a comprehensive framework for virtual-real integration centered on unmanned systems, aiming to unify fragmented research efforts and guide future exploration. By adopting both global and local perspectives, the framework facilitates a cohesive understanding of system components, their interactions, and the structural relationships that underpin seamless coordination. This is further complemented by an application case study and a synthesis of the current limitations. To operationalize this framework, the paper systematically examines enabling technologies, existing constraints, and the collaborative architecture that supports dynamic interaction across physical and virtual domains. It further outlines the evolving trends and construction requirements of practical application scenarios. Finally, key challenges and emerging research opportunities are discussed to inform future work and encourage deeper exploration of this rapidly developing field. Note to Practitioners-The motivation of this work is to explore advancement opportunities for unmanned systems in the trend of fusion between virtual and real worlds, so as to facilitate their better deployment in the real world. The impact of relevant enabling technologies and restrictions, the enhancement of unmanned systems, the involvement of human intelligence, and the deployment of specific application scenarios are all crucial for researchers and practitioners in the field to carry out practical work. This paper can provide researchers and practitioners with a comprehensive reference including the above content and help them utilize the integration of virtual and real environments to enhance the actual deployment of unmanned systems by further focusing on virtual application scenarios. In addition, challenges and future research directions identified in this survey also help identify entry points driving developments in the field.

Embodied agents represent a prominent research focus across both academia and industry. The prevailing paradigm has gradually shifted from reactive assistance, which requires explicit user queries, to proactive assistance, capable of recognizing human needs and offering support without explicit instructions. Nevertheless, existing studies on proactive assistance remain confined to narrow scenarios and primarily emphasize task completeness, whereas real-world agents must operate in open-domain environments while adhering to social expectations. To bridge this gap, we extend the concept of proactive assistance to Social Proactive Intelligence (SPI), characterized by diverse scenarios, social understanding, and robot-centric behaviors. We further introduce RobotEQ, a dedicated benchmark for SPI. We first define two tasks: behavior judgment, emphasizing global contextual understanding, and spatial grounding, focusing on local perceptual details. Building on these tasks, we construct RobotEQ-Data, a dataset comprising 1,812 synthetic and 223 real-world scenarios, 7 social facets, 22K+ human annotations, 3K+ behavior judgment questions, and 3K+ spatial grounding questions. Furthermore, we establish RobotEQ-Bench to evaluate the performance of representative models. Experimental results demonstrate that current models fall short of achieving reliable SPI. Further analysis reveals that incorporating external social knowledge yields consistent improvements. This work aims to advance the development of socially desirable embodied agents in open-domain environments.

The robust manipulation of deformable objects (DOs), such as cables, clothes, and food, is essential for developing next-generation robotic systems in industrial, service, and healthcare applications. However, achieving a reliable system for these tasks has been historically challenging. Unlike rigid objects, the infinite-dimensional state space, severe self-occlusions, and complex dynamics of DOs present significant barriers to robotic perception, modeling, and manipulation. The development of data-driven learning and the recent emergence of foundation models have enabled novel techniques for deformable object manipulation (DOM). These techniques, based on data-driven paradigms, can address some of the challenges that analytical approaches in DOM face. However, some existing reviews do not include all aspects of DOM, and some previous reviews do not summarize data-driven approaches adequately. In this article, we survey more than 150 relevant studies and summarize recent advances, open challenges, and new frontiers in perception, modeling, and manipulation of DOs. We regard research in DOM as paving the way toward general-purpose robotic systems, and we outline key future directions for generalizable robotic manipulation. Specifically, we advocate for the algorithmic synergy of vision-language-action models with reinforcement learning and world models, a universal high-DoF dexterous hand with tactile sensors, and comprehensive evaluation at scale.

Out-of-distribution (OOD) inputs, which deviate from a model's training distribution, are a central obstacle to the reliable deployment of learning-based models, and in humanoid robot control this becomes a direct motion-execution risk. Reinforcement-learning-based imitation control enables teleoperation, motion replay, and whole-body skills, but reliably tracks only commands close to its controller-training motion distribution. At deployment, open-ended commands fall OOD through retargeting artifacts, generative discontinuities, or inconsistent teleoperation, and executing them causes tracking divergence and imitation failure.We propose a diffusion-based Motion Filter that protects a pretrained imitation controller by scoring the OOD deviation of incoming commands and replacing high-risk ones. It rests on a shared autoregressive diffusion prior trained on controller-training motion data, whose perturb-denoise reconstruction gives an initial measure of each command's deviation. A temporal LSTM then learns a predictive mapping from this deviation to imitation risk, letting Motion Risk Identification flag the risky segments. The same prior then generates in-distribution replacements, with reinforcement-learning guidance selecting the initial noise.We evaluate the filter in Isaac Sim on an H1-2 humanoid. Under a fixed temporal classifier, perturb-denoise diffusion scoring raises segment-level F1 by 32% over the best non-diffusion scorer, achieving the strongest temporally coherent risk detection among all compared unsupervised methods. Acting on these estimates, the filter cuts cumulative instability 52.8x (0.264 to 0.005) and reduces terminations from 165 to 2, with a smoother return to the input stream. The filtered commands are confirmed to remain executable on a physical humanoid through hardware experiments.

Sustainable energy supply and high-speed communications are two significant needs for the upcoming 6G applications. This paper introduces a self-aligning resonant beam system for simultaneous light information and power transfer (SLIPT), employing a novel coupled spatially distributed resonator (CSDR). The system utilizes a resonant beam for efficient power delivery and a second-harmonic beam for concurrent data transmission, inherently minimizing echo interference and enabling bidirectional communication. Through comprehensive analyses, we investigate the CSDR's stable region, beam evolution, and power characteristics in relation to working distance and device parameters. Numerical simulations validate the CSDR-SLIPT system's feasibility by identifying a stable beam waist location for achieving accurate mode-match coupling between two spatially distributed resonant cavities and demonstrating its operational range and efficient power delivery across varying distances. The research reveals the system's benefits in terms of both safety and energy transmission efficiency. We also demonstrate the trade-off among the reflectivities of the cavity mirrors in the CSDR. Besides, an experiment was conducted to verify the feasibility of self-aligning beam generation and safety under the designed structure. These findings offer valuable design insights for resonant beam systems, advancing SLIPT with significant potential for remote device connectivity.

The ship system is a complex hull-propulsion-shaft-power multibody system composed of at least one shaft subsystem coupling with numerous flexible bodies. To achieve rapid and precise modeling of the ship system, this paper combines the advantages of rapid modeling and efficient simulation of the Multibody System Transfer Matrix Method (MSTMM) for the shaft system with the power of the Finite Element Method (FEM) for flexible bodies. The fundamental principles of the Finite-element Multibody System Transfer Matrix Method (FMSTMM) are elaborated. The general multibody system comprising the shaft and hull is divided into subsystems using MSTMM and subsystems using FEM and modal synthesis method. The dynamic equations of each subsystem and the overall transfer equation of the entire system are derived separately. The constraint relationships at the subsystem connection points are utilized to formulate the overall transfer equation and characteristic equation of the multibody system, in order to compute of the system’s dynamic characteristics. A FMSTMM dynamic model and algorithm for the ship multibody system including shaft and hull subsystems are established. The efficient modeling and simulation of the vibrational characteristics of the shaft-hull system are achieved. The simulation and test results meet well, validating the proposed method.

Underground environments, including tunnels, mines, and collapsed structures, impose severe constraints on Unmanned Aerial Vehicle (UAV) communication because of signal attenuation, Non-Line-of-Sight (NLoS) propagation, intermittent connectivity, and Global Positioning System (GPS)-denied operation. Conventional communication and networking models, which are often designed for open-air or infrastructure-supported environments, are difficult to adapt to these dynamic, partially observable, and data-sparse conditions. Digital Twin (DT) technology offers a promising direction by enabling virtual representations of physical UAV systems and underground communication environments for monitoring, simulation, prediction, and decision support. When combined with Artificial Intelligence (AI), DTs can support predictive channel modeling, synchronization under incomplete feedback, adaptive communication-mode selection, and distributed learning across UAV networks. However, existing surveys have generally examined UAV communication, DT technologies, underground systems, or AI-enabled networking as separate research areas, leaving the intersection of AI-enabled DTs and underground UAV communication insufficiently systematized. To address this gap, this paper presents a structured survey of AI-enabled DT approaches for underground UAV communication. It organizes the literature according to DT objectives, communication modalities, synchronization mechanisms, AI integration levels, and mission-specific Quality-of-Service (QoS) requirements. The paper further develops a conceptual architecture that links the physical underground UAV environment, cyber twin layer, intelligence layer, and application-control layer. In addition, it analyzes the roles of predictive AI, DT-assisted synchronization, Explainable AI (XAI), and Federated Learning (FL) in improving communication reliability, operational transparency, and swarm-level adaptability under subterranean constraints. Finally, the paper identifies open challenges related to channel modeling, dataset scarcity, synchronization fidelity, energy-efficient intelligence, interoperability, security, and real world validation, and presents a research roadmap for developing robust, trustworthy, and adaptive DT-enabled underground UAV communication systems.

The computational intelligence of robots has advanced significantly in vision recognition and natural language processing. Yet robots are still struggling with tasks involving physical interactions. Although efforts have been made in their haptic perception, the interpretation of tactile sensory information for robots is lacking. Here, based on cognitive developmental robotics, we integrate tactile sensor signals with robot proprioception to estimate the space-force boundary, which extends the robot’s spatial intelligence with force to enhance its embodied capabilities in scenarios such as human-robot close physical interaction. To this end, drawing lessons from human haptic feeling and infant learning, a meta-learning-based neural framework is proposed, including the fibring-paradigm-based neural network, inductive-inference-based training process, and active-learning-based input construction, with competitive generalization, resilience, and scalability. It can provide robots with a task-agnostic practical clue for physical interaction, including contact properties on interfaces and contact meanings on robot bodies, from data of robot actions and tactile sensors only, without sophisticated robot models. It also has potential for continuous optimization through few-shot and few-epoch learning based on new physical interactions. Furthermore, its general application effectiveness is demonstrated in several human-robot physical interaction tasks in the real world.

Mechanical forces are emerging physical cues that regulate biochemical signals of immune cells for antitumor immunity. Owing to the lack of precise tools to impose intracellular forces, little is known about whether and how organelle-level forces trigger mechanotransduction for antitumor immunity. Here, we developed a magneto-mechanical force-triggered lysosomal membrane permeabilization (MagLMP) strategy to induce durable macrophage repolarization for in vivo applications. Self-assembled magnetic nanomotors are driven by rotational magnetic fields, facilitating dynamic damage to the lysosomal membrane by a finely tuned torque-induced vortex. Intriguingly, galectin 9 (Gal9) was found to be critical for sensing cyclic MagLMP, which dynamically activated AMP-activated protein kinase (AMPK), enhanced activation of nuclear factor kappa B (NF-κB), and induced metabolic alterations for sustained M1-like macrophage repolarization, followed by mounting of antitumor immunity. Through single-cell RNA sequencing of tumor tissues, as well as macrophage depletion-reconstitution models involving intratumoral transfer of Gal9-KO bone marrow-derived macrophages (BMDMs) and AMPK shRNA-transduced Gal9-KO BMDMs, we confirmed the Gal9-AMPK-NF-κB axis as the essential pathway by which MagLMP functions in antitumor therapy. In a mouse model of lung adenocarcinoma in situ, overall survival was extended after intravenous administration of nanomotors followed by cyclic MagLMP, and one third of mice survived for more than 300 days. Together, these results demonstrate an intracellular mechanical strategy that can dynamically manipulate innate immune responses in vivo, providing a tool for durable immunotherapy through organelle mechanotransduction.

Building scalable and reusable multi-agent decision policies from offline datasets remains a challenge in offline multi-agent reinforcement learning (MARL), as existing methods often rely on fixed observation formats and action spaces that limit generalization. In contrast, large language models (LLMs) offer a flexible modeling interface that can naturally accommodate heterogeneous observations and actions. Motivated by this, we propose the Decision Language Model (DLM), which formulates multi-agent decision making as a dialogue-style sequence prediction problem under the centralized training with decentralized execution paradigm. DLM is trained in two stages: a supervised fine-tuning phase, which leverages dialogue-style datasets for centralized training with inter-agent context and generates executable actions from offline trajectories, followed by a group relative policy optimization phase to enhance robustness to out-of-distribution actions through lightweight reward functions. Experiments on multiple benchmarks show that a unified DLM outperforms strong offline MARL baselines and LLM-based conversational decision-making methods, while demonstrating strong zero-shot generalization to unseen scenarios across tasks.