Artificial intelligence (AI) is undergoing a structural transition from isolated digital assistance to organized closed-loop systems that can perceive, reason, act, and evolve. While large-scale models and AI agents have expanded the capabilities of digital intelligence, the next stage of AI will be increasingly defined by its connection to physical action, parallel simulation, organizational coordination, and continuous learning. Physical intelligence enables embodied agents to sense and act in real environments, whereas parallel intelligence provides a computational space for representing, simulating, and optimizing real systems before decisions return to physical execution. This editorial argues that the future of AI should be understood as an organized closed-loop paradigm in which physical intelligence, parallel intelligence, and organizational intelligence reinforce one another. Through intelligent organizations and AI education as representative cases, we show how external demands, agents, workflows, knowledge, resources, and feedback can be integrated into adaptive operating systems. Under this framework, intelligence is no longer merely a tool used by organizations, but becomes the mechanism through which organizations sense, coordinate, decide, act, learn, and evolve.

Automated cooking robots have traditionally relied on predefined procedures and rule-based control, ensuring stable execution but offering limited personalization, whereas recent large-model approaches support natural language interaction but often suffer from opaque decision making and unreliable execution in real kitchens. To address this challenge, this paper proposes an agentic framework that systematically decomposes personalized cooking requirements into structured and verifiable control programs rather than directly mapping language to actions. Multiple artificial intelligence (AI) agents collaboratively transform user intents into canonical recipes, workflow programs with explicit flow control, and executable Python code grounded in an atomic action library. The system consists of three tightly coupled stages: offline recipe-to-code generation through multiple agents, online closed-loop execution with supervisory intervention enabled by multimodal perception, and post-run adaptation that updates user preference models for long-term personalization. Real-world experiments on a physical cooking platform demonstrate that the proposed framework achieves reliable task completion, transparent execution logic, and effective anomaly handling across diverse personalized scenarios, validating its practicality for trustworthy automated cooking in real environments.

To address the challenges in solar greenhouses where cucumber canopy point clouds are highly dense, leaves exhibit severe overlap, and inter-plant adhesion is pronounced, resulting in insufficient accuracy and poor robustness of traditional point cloud segmentation algorithms, this study proposes a comprehensive cucumber hierarchical segmentation framework (Cuc-HiSeg) for terrestrial laser scanning (TLS) point clouds, enabling high-precision segmentation from population level to individual plants and further to leaves. High-resolution three-dimensional (3D) point cloud data of cucumber canopies at the seedling, early flowering, and fruiting stages were collected using TLS. For segmentation from population to individual plants, an improved hierarchical density-based spatial clustering of applications with noise (HDBSCAN) algorithm integrating k-dimensional tree (KD-tree) optimization and secondary clustering with mean shift was developed, improving the F1-score by 0.7%, 5.0%, and 8.0% across the three growth stages compared with the original algorithm. For leaf segmentation, a leaf point cloud data augmentation algorithm named CuLeafExpander was proposed to enhance the generalization capability of deep learning models. Experimental results demonstrate that the improved HDBSCAN and region-growing (RG) algorithms attained F1-scores above 91.9% and 93.9%, respectively. The proposed Cuc-HiSeg provides technical support for efficient processing of crop point clouds and advances digital crop phenotyping and intelligent greenhouse management.

This paper investigates the optimal control of discrete-time nonlinear systems with two strategically coupled control inputs arising from a Nash-type interaction. By augmenting the coupled inputs into a unified control vector and embedding the equilibrium conditions into a single performance index, the original game-based problem is reformulated as an augmented optimal control problem. Under standard regularity assumptions, we establish that the Nash stationarity conditions are locally equivalent to the first-order optimality conditions characterized by the Pontryagin maximum principle (PMP). Based on this reformulation, two PMP-based iterative algorithms are developed to solve the resulting forward-backward difference equations (FBDEs). The first algorithm exploits exact second-order information through the inversion of a regularized Hessian matrix, while the second constructs a Hessian-free recursive update that avoids explicit Hessian inversion. For both methods, a rigorous local convergence analysis is provided, showing that the effective linear contraction factor vanishes as the iteration horizon increases, leading to a superlinear-type local convergence behavior. Numerical experiments on representative nonlinear systems validate the theoretical findings and demonstrate improved convergence efficiency compared with classical gradient-based and Newton-type methods. The proposed theoretical framework and algorithms are particularly relevant to cooperative control problems in unmanned aerial vehicle (UAV) swarms, where multiple decision-makers interact through coupled dynamics and performance objectives.

Highway bottlenecks caused by temporary lane closures can lead to severe congestion and capacity drop, especially under high traffic demand. This paper proposes a learning-based Lagrangian traffic control framework in which a small fraction of connected and automated vehicles (CAVs) in mixed traffic are organized into specially structured multi-lane fleets and utilized as mobile actuators to mitigate congestion. A hierarchical control architecture is developed, where an upper-level reinforcement learning (RL) agent determines fleet-level control parameters, including cruising speed and inter-vehicle spacing, while a lower-level model predictive control (MPC) module ensures real-time trajectory tracking and safety. The proposed method explicitly captures cross-lane CAV cooperation within fleets through the design of the RL action space, and incorporates CAV-induced traffic regulation effects into the reward formulation to optimize traffic dynamics and alleviate bottleneck congestion. Simulation results in a multi-lane highway scenario demonstrate that the proposed strategy significantly improves traffic performance across a wide range of demand conditions, reducing travel delay and congestion length while increasing average traffic speed. In addition, the method exhibits generalization across varying traffic demands and improved performance with increased availability of CAV fleets.

This paper investigates bipartite-scaled consensus for leader-follower linear multi-agent systems (MASs) under directed fixed or switching cooperative-competitive topologies. The followers are partitioned into two subgroups with cooperative interactions within each subgroup and competitive interactions across subgroups. Prescribed scaling factors are introduced to achieve specified proportional relationships, thereby relaxing the conventional bipartite consensus condition that forces the two subgroups to converge to equal magnitude but opposite sign. This feature is particularly beneficial for complex systems where different subgroups require distinct influence levels from the leader. When full state information is available, a distributed state feedback controller is developed and consensus conditions are derived via algebraic graph theory and Lyapunov stability theory. For scenarios where states are not directly measurable, an observer-based output feedback controller is constructed and the closed-loop stability is rigorously analyzed. Furthermore, both strategies are extended to switching topologies with corresponding sufficient conditions established. Numerical simulations validate the theoretical findings.

The vehicle to everything (V2X) communication of intelligent and connected vehicles faces uncertainties such as random latency and packet loss. These issues pose safety challenges. This paper proposes a closed-loop integrated strategy that combines perception, decision-making, and control. A communication uncertainty metric is established based on recent packet loss rate and average delay. A three-state machine (cooperative cruise mode, perception uncertainty mode, and local safety mode) is designed to adjust the control strategy based on a preset uncertainty threshold. A proportion integration differentiation (PID) controller ensures safe car following. MATLAB simulations under high uncertainty show that the proposed method guarantees that the following distance always satisfies the dynamic safety criterion (a time headway of 1.5 s plus a standstill gap), whereas the fixed-strategy baseline fails under high uncertainty, significantly outperforming a fixed-strategy baseline. This method is computationally efficient and can be deployed on embedded devices. Simulation results validate the effectiveness of this approach in handling V2X communication uncertainties.

Human-agent interaction (HAI) is an increasingly important research area that investigates how humans interact with intelligent agents. Effective HAI relies heavily on social signal understanding, particularly visual social signals that convey human attention, intention, and affect, enabling agents to perceive, interpret, and respond appropriately to human behaviour. Recent advances in artificial intelligence (AI) have substantially improved visual perception and representation learning, allowing machines to detect and model visual social behavioural cues with increasing accuracy. However, translating these cues into socially meaningful interpretations remains a significant challenge, often resulting in unnatural interactions and limited empathic responsiveness in HAI systems. In this paper, we coined the concept of computational social vision (CSV) spanning the domain of both psychology and computing science to understand social signals. We summarize visual-cognitive computing techniques for this purpose with an emphasis on how social vision and artificial intelligence technologies support the inference of latent social states from dynamic behavioural evidence. To structure the field, we organise representative methods according to the semantic roles of visual behaviours, including both head-centric and body-centric social vision approaches. Finally, we discuss socially interactive embodied systems, analyzing how agents, such as social robots and virtual humans, adaptively perceive their environments and interact with humans as well as the future trend.

The development of reliable autonomous driving systems for mining is critically limited by the scarcity of long-tail scenario data. While generative models provide a promising direction, they often lack the fine-grained control and domain-specific realism required in safety-critical applications. To address this challenge, we propose a decoupled training pipeline that models a scene's fundamental components, namely style, layout, and motion, as architecturally separated and controllable factors. Firstly, style and layout are jointly fine-tuned to capture the unique textures and spatial structures of mining environments, ensuring high-fidelity visuals and precise spatial control. Subsequently, we selectively fine-tune the temporal layers of the video generation model. This captures empirically plausible visual motion patterns of heavy machinery, creating short, high-fidelity videos. Experiments show that our pipeline not only outperforms general video baselines but also improves downstream task performance. By bridging the gap between static and dynamic data, our framework supports robust perception and supplies synthetic dynamic scenarios for vision-language-model (VLM)-based training and evaluation in autonomous mining.

As an ideal clean fusion fuel, helium-3 (³He) holds broad application prospects. However, its terrestrial reserves are extremely scarce, whereas lunar regolith contains abundant ³He, making large-scale in-situ mining crucial. Because it exists in extremely low concentrations and is primarily enriched within the surface lattices of minerals, such as fine-grained ilmenite, extracting ³He requires a complex and multi-disciplinary engineering system involving physical excavation and thermal processing. This paper presents an architecture for large-scale in-situ ³He mining that encompasses four main stages: excavation, sorting, extraction, and purification. It also defines the collaborative operational workflows among these stages. Building on this framework, this paper introduces an integrated lunar regolith collection vehicle capable of large-scale excavation and sorting, along-side a detailed process flow for ³He extraction and purification. The mining architecture and process flows proposed in this paper establish a theoretical foundation and technical reference for the engineering implementation of future in-situ lunar ³He mining.

Physical artificial intelligence (AI) is becoming an important paradigm for embodied intelligence, where agents must reason, imagine, and act under real-world physical constraints. However, current physical reasoning model (PRM) and physical generation model (PGM) may still produce physically inconsistent judgments, scenes, and motions, exposing the lack of a verifiable truth anchor for reliable physical intelligence. Existing simulation-based methods can provide local physical supervision, but their limited coverage makes them insufficient for open and evolving physical environments. To address this bottleneck, this letter proposes parallel physical intelligence (PPI), a truth-anchored co-evolution framework based on parallel intelligence and the artificial systems, computational experiments, and parallel execution (ACP) methodology. PPI introduces the physical fidelity anchor (PFA) to calibrate both PRM and PGM, while enabling PRM and PGM to co-evolve through mutual data augmentation and physical plausibility feedback. During parallel execution, embodied agents and disembodied agents interact through a virtual-real closed loop, where real-world failures and novel physical phenomena update the PFA and trigger subsequent co-evolution. This provides a scalable and adaptive pathway for physically grounded embodied intelligence in complex environments.

Scientific knowledge is increasingly becoming part of the operational substrate of intelligent systems and intelligent industries, rather than serving only as a record of scholarly communication. As these systems move toward closed-loop integration of perception, decision, and control, the infrastructures that produce, validate, update, and circulate knowledge become correspondingly important. Yet contemporary scientific publishing still relies on static articles, delayed review cycles, document-centered dissemination, and weak machine readability, creating a knowledge-automation bottleneck for agent-based intelligence. Drawing on parallel intelligence and the artificial systems, computational experiments, and parallel execution (ACP) approach, this letter proposes parallel publishing: a living knowledge infrastructure that runs in parallel with physical, cyber, and social systems, continuously transforming research signals, operational feedback, and scientific claims into structured, verifiable, and machine-usable knowledge units. A three-layer architecture is presented, together with a conceptual representation of structured knowledge units and four defining design properties: continuity, structured verifiability, incentive alignment, and decentralized resilience. An autonomous-driving scenario further illustrates how operational feedback can be converted into verifiable knowledge for intelligent systems. The analysis suggests that closing the intelligence loop depends not only on advances in models, sensors, and hardware, but also on the emergence of a knowledge layer that is current, auditable, and structured for machine use across intelligent industries and societies.