ai agents

Autonomous robotic manipulation critically requires world models capable of understanding the physical mechanisms and dynamics of the environment; however, existing definitions remain ambiguous and capability boundaries ill-defined, hindering both generality and practical deployment. Method: This work adopts a functional perspective to systematically characterize the core capabilities of world models for robotic manipulation, proposing a task-driven, unified framework that integrates state representation learning, dynamic modeling, sequence prediction, closed-loop control, and model-based reinforcement learning. Contribution: We formally identify the essential components and functional roles of world models within the perception–prediction–control loop for the first time. Moving beyond conventional static representations, we emphasize dynamic modeling, causal reasoning, and online planning as critical capabilities. Furthermore, we provide a clear capability taxonomy and a systematic construction methodology, enabling the development of generalizable, scalable world models for robotics.

This work systematically investigates safety risks of world models in embodied intelligent agents, specifically within autonomous driving and robotics. Addressing the potential for hazardous environmental or agent-level consequences arising from inaccurate predictions and unsafe control generation, we propose the first safety risk analysis framework tailored to embodied world modeling. Our methodology integrates literature review with empirical evaluation: we collect prediction outputs from state-of-the-art world models, identify and categorize recurrent failure modes—including temporal misalignment, physical inconsistency, and boundary violation—and establish a pathology-driven taxonomy. We further design quantitative metrics for assessing predictive safety and validate their efficacy across multiple case studies. The study uncovers critical vulnerabilities in current world models, revealing fundamental trade-offs between prediction fidelity and safety guarantees. These findings provide both theoretical foundations and empirical evidence to guide the development of safety-aware world modeling techniques.

Autonomous AI systems struggle to simultaneously satisfy procedural, legal, and ethical constraints in real-world environments, leading to decision-making dilemmas under conflicting norms. Method: We propose a dynamic decision-making framework that integrates normative, pragmatic, and contextual knowledge. It enables AI to autonomously generate candidate actions under constraint conflicts, evaluate them across multiple dimensions—including goal consistency and value alignment—and produce human-interpretable action justifications. Technically, the framework unifies multi-source knowledge reasoning, fine-grained situational understanding, and explainable planning—moving beyond limitations of end-to-end policy learning. Contribution/Results: This work is the first to systematically characterize the knowledge types and mechanisms required for compliant and reasonable decision-making in underspecified scenarios. Empirical evaluation demonstrates substantial improvements in behavioral planning robustness, contextual adaptability, and alignment with human values.

This work addresses the challenge of autonomous goal setting and generalizable learning in reinforcement learning (RL) without external rewards. Methodologically, it reformulates standard RL environments into goal-conditioned ones and introduces an environment-agnostic, self-supervised goal-generation mechanism that requires no extrinsic reward and supports arbitrary observations as goals, while remaining compatible with off-policy RL algorithms. Its core contribution lies in decoupling goal generation from policy learning, enabling stable, reward-free exploration. Empirically, the approach achieves significantly higher average goal success rates compared to conventional methods—without increasing training time—and demonstrates strong cross-environment generalization across diverse, heterogeneous domains. These results validate both its environment independence and practical efficacy for unsupervised, goal-directed skill acquisition.

This paper addresses adaptive decision-making for autonomous agents in unknown environments. Methodologically, it introduces a unified framework integrating model-based planning and model-free reactive control: (1) Meta-Interpretive Learning (MIL) induces interpretable, plannable logical rules from task trajectories to construct a symbolic model-based solver; (2) this solver serves as a supervisory signal to train a model-agnostic controller that requires no prior environmental knowledge. The core contribution is the first demonstration of strict equivalence in navigation-solving capability between the two paradigms: across randomized mazes and open-water lake environments, the controller perfectly replicates all tasks solved by the symbolic solver. This establishes the feasibility of synergistically combining symbolic reasoning with end-to-end learning, thereby providing a novel pathway toward explainable, highly adaptive intelligent agents.

Real-time multi-step planning and obstacle avoidance for autonomous robots in dynamic environments remain challenging, particularly under resource constraints and without prior map knowledge. Method: We propose a lightweight, closed-loop reactive planning framework that requires no pre-mapping or offline computation. Our approach integrates biologically inspired attention mechanisms with local LiDAR perception to construct transient control-chain plans. It introduces forward depth-first model checking—novel in real-time multi-step planning—combined with environment-aware 2D LiDAR discretization and closed-loop feedback control. Contribution/Results: The framework provides theoretical guarantees on safety and interpretability. Empirically, it generates safe, multi-step local trajectories within 100 ms on low-power embedded hardware. In complex scenarios—including dead ends and playgrounds—it significantly outperforms single-step reactive systems in obstacle avoidance success rate and response robustness.

This paper addresses the foundational question: *When can a physical system be rigorously regarded as an agent possessing beliefs and goals?* We propose a POMDP-based explanatory framework, formalizing an agent as a physical system satisfying two joint constraints: (i) its state evolution must conform to Bayesian belief-updating dynamics, and (ii) its policy must be optimal with respect to a specified objective. Crucially, we introduce the *completeness of a POMDP solution*—simultaneous adherence to correct belief evolution and optimal action selection—as a necessary and empirically falsifiable criterion for agency, overcoming the limitations of prior definitions relying solely on belief mapping. This yields the first axiomatization of agency that is mathematically rigorous, computationally operational, and empirically testable. It reveals necessary constraints linking physical dynamics to agential properties, thereby establishing a theoretical foundation for AI safety verification and computational modeling of consciousness.

This work addresses the tendency of large language model agents to prematurely rely on prior knowledge in unfamiliar environments, leading to inadequate exploration and task failure. The authors propose an Explore-then-Act paradigm that decouples exploration from execution: agents first systematically gather environmental information within a fixed interaction budget, then leverage the acquired embodied knowledge to accomplish tasks. The study formally characterizes an agent’s autonomous exploration capability, introduces a verifiable exploration coverage metric, and devises an alternating training strategy that interleaves exploration and task objectives. Furthermore, a dual-trajectory reinforcement learning framework with a verifiable reward mechanism is introduced to optimize behavior policies. This approach substantially enhances generalization in unseen environments, overcoming the limitations of conventional methods whose narrow behavioral repertoires constrain downstream task performance.

This work addresses two key challenges in multi-agent systems: (i) difficulty in inferring non-cooperative agents’ intentions without direct observation, and (ii) difficulty in achieving safety-critical decision-making under resource constraints. Methodologically, we propose a robust interaction framework integrating perception modeling and risk-aware decision-making: (i) a latent-intent estimation model grounded in historical trajectories enables behavior prediction without explicit intent signals; (ii) a dual-path decision mechanism jointly performs state estimation, behavior prediction, and risk analysis—integrating reinforcement learning and game-theoretic reasoning. Our key contribution is the first integration of risk-aware control into an intent-driven behavioral prediction feedback loop, ensuring provable safety-performance trade-offs under computational constraints. Experiments demonstrate significant improvements in prediction accuracy and interaction robustness under high uncertainty, validating the framework’s efficacy for safety-critical applications such as power systems and autonomous driving.

This work addresses the challenge of enabling agents to implicitly convey internal state information through their actions in communication-constrained environments, thereby facilitating accurate external observation. The authors propose a method that directly embeds state observability into the reinforcement learning reward function, guiding the policy to actively expose informative state signals while preserving primary task performance. By integrating reinforcement learning with observability-aware optimization, the approach successfully trains control policies with high observability in an aerial tracking task. Experimental results demonstrate that the resulting policies significantly enhance the accuracy of state reconstruction by external observers, with negligible degradation to the main task performance.

Traditional large language models (LLMs) lack autonomy and sustained interactive capability, limiting their effectiveness in complex real-world tasks. To address this, we propose an agentive LLM architecture that integrates environment perception, multi-step reasoning (Chain-of-Thought and Tree-of-Thought), hierarchical short- and long-term memory mechanisms, and action execution interfaces—enabling end-to-end autonomous decision-making loops. Unlike passive, reactive LLMs, our architecture supports proactive goal decomposition, dynamic state maintenance, and adaptive behavior driven by environmental feedback. Experimental results demonstrate substantial improvements in task completion rates and generalization across diverse benchmarks, significantly narrowing the performance gap with human baselines. The framework provides a scalable, unified foundation for advancing LLMs toward embodied, persistent agency.

This work proposes a decentralized multi-agent reinforcement learning (MARL) approach to address the challenges of communication constraints, dynamic obstacles, and partial observability in GNSS-denied indoor environments for collaborative multi-drone exploration. Implemented on the high-fidelity Godot simulation platform, the method integrates LiDAR-based perception with local occupancy map sharing and models the problem as a networked distributed POMDP (ND-POMDP) to enable communication-aware cooperative exploration in continuous action spaces. By abandoning conventional reliance on discrete actions, centralized control, prior maps, and persistent connectivity, the approach introduces curriculum learning and a lightweight neural architecture, significantly enhancing training efficiency, robustness, and scalability. This provides a practical and efficient solution for real-world deployment of multi-drone systems.