Vision-language-model (VLM)-guided reinforcement learning (RL) has recently attracted significant attention for it, replacing brittle hand-crafted rewards with semantically grounded signals; however, deploying such simulation-trained policies on real vehicles remains a fundamental challenge, because they rely on simulator-native observations and simulator-coupled action semantics with no counterpart on physical hardware. We identify a general principle: the simulation-to-reality gap decomposes into two largely orthogonal axes, a sensing-and-dynamics domain gap and a task-and-geometry gap, the former closable without real-world policy training by re-projecting real perception and control onto the policy's training manifold. We formalize this as a transfer guarantee that bounds the deployment gap by three independently controllable error terms, and instantiate it as Sim2Real-AD, which combines a Geometric Observation Bridge, a Physics-Aware Action Mapping, a Two-Phase Progressive Training curriculum, and a Real-time Deployment Pipeline. As a proof of concept, a CARLA-trained VLM-guided RL policy is transferred zero-shot to a full-scale battery-electric Ford E-Transit van in Madison, WI, USA, and drives across car-following, obstacle-avoidance, and stop-sign scenarios using no real-world training data. To our knowledge, this is among the first zero-shot closed-loop deployments of a CARLA-trained VLM-guided RL policy on a full-scale real vehicle, and the decomposition offers a principled, broadly applicable route for moving simulation-trained, foundation-model-guided policies into the physical world, supporting energy-efficient intelligent driving on electrified transportation platforms. The demo video, code, and model checkpoint are available at: https://zilin-huang.github.io/Sim2Real-AD-website/.
When individual robots have limited sensing capabilities or insufficient fault tolerance, it becomes necessary for multiple robots to form teams during exploration, thereby increasing the collective observation range and reliability. Traditionally, swarm formation has often been managed by a central controller; however, from the perspectives of robustness and flexibility, it is preferable for the swarm to operate autonomously even in the absence of centralized control. In addition, the determination of exploration targets for each team is crucial for efficient exploration in such multi-team exploration scenarios. This study therefore proposes an exploration method that combines (1) an algorithm for self-organization, enabling the autonomous and dynamic formation of multiple teams, and (2) an algorithm that allows each team to autonomously determine its next exploration target (destination). In particular, for (2), this study explores a novel strategy based on large language models (LLMs), while classical frontier-based methods and deep reinforcement learning approaches have been widely studied. The effectiveness of the proposed method was validated through simulations involving tens to hundreds of robots.
This article addresses the challenging problem of constrained optimal control for human–machine systems subject to external disturbances and the bounded rationality of the human operator. To this end, a novel game-theoretic framework is proposed. Unlike monolithic game formulations, the framework uniquely disaggregates the control problem by transforming it into a multifaceted game via logarithmic barrier functions (BFs): it models human–machine cooperation as a positive-sum game oriented toward shared objectives, and disturbance rejection as a zero-sum game tailored for robustness enhancement. To capture the nonideal human decision-making, we integrate the level-$k$reasoning framework to model the operator’s bounded cognitive dynamics. The corresponding coupled Hamilton–Jacobi–Isaacs (HJI) equations for this human–machine game are derived, and critically, a rigorous proof of global asymptotic stability (GAS) for the transformed system is provided, establishing a solid theoretical foundation. For online implementation without requiring prior knowledge of the system dynamics, we develop a resource-efficient learning architecture based on the adaptive dynamic programming (ADP) and a novel dynamic event-triggered mechanism (DETM). A key feature of this architecture is a fuzzy logic-based module for dynamic authority allocation, which adaptively adjusts control sharing in real time. Rigorous analysis demonstrates that all signals in the closed-loop system are uniformly ultimately bounded and that Zeno behavior is precluded. Simulation results are presented to validate the effectiveness and superiority of the proposed control strategy.
The Cyber-Physical Systems (CPS) experience great operational complexity in the stochastic and non-stationary contexts when the control logic can be considered static. The current paper introduces an autonomous agent-based intelligence framework, which can be used in perpetually adaptive and decentralized organization of heterogeneous CPS systems. The given framework implements a multi-agent system (MAS) to bridge the gap between the cyber and physical layers with the help of built-in perception-action loops and planning modules. With the help of online reinforcement learning (RL) and predictive analytics, individual agents change control policies in real-time to alleviate disturbances in a system and changing operational constraints. A shared knowledge layer that is distributed is put in place to coordinate inter-agent coordination and policy refinement making the system-wide scalable and fault-tolerant. Experimental validation in a wide range of CPS situations proves that the suggested autonomous structure can be much more efficient and self-optimizing than centralized baseline frameworks, providing a stable channel of self-evolutionary industrial and robotic systems.
In this work, we present a novel machine learning framework called Self-Directed Task Identification (SDTI), which enables models to autonomously identify the correct target variable for each dataset in a zero-shot setting without pre-training. SDTI is a minimal, interpretable framework demonstrating the feasibility of repurposing core machine learning concepts for a novel task structure. To our knowledge, no existing architectures have demonstrated this ability. Traditional approaches lack this capability, leaving data annotation as a time-consuming process that relies heavily on human effort. Using only standard neural network components, we show that SDTI can be achieved through appropriate problem formulation and architectural design. We evaluate the proposed framework on a range of benchmark tasks and demonstrate its effectiveness in reliably identifying the ground truth out of a set of potential target variables. SDTI outperformed baseline architectures by 14% in F1 score on synthetic task identification benchmarks. These proof-of-concept experiments highlight the future potential of SDTI to reduce dependence on manual annotation and to enhance the scalability of autonomous learning systems in real-world applications.
In this paper, I prove that sublinear regret across the environment Class C requires six functional properties, that these properties are mutually independent, and that they compose into a directed informational chain closing back on itself — a six-link cycle whose final link is grounded in an explicit Doob martingale construction over cycles of play. All six properties are defined functionally — as conditions on the distributions a decision-maker induces over actions and canonical summaries — so the results are invariant under implementation and apply to any decision-making system that can be modelled within the class: a person, an institution, or a machine. Every theorem in this paper, without exception, is checked line by line in the Lean 4 proof assistant against Mathlib: the formalization (~12,700 lines) contains zero `sorry`, zero custom axioms, and zero opaque definitions. Class C is the union of all POMDPs satisfying at least one of six structural properties covering the fundamental qualitative dimensions of adaptive hardness: reward ambiguity (P1), absorbing traps (P2), local optima (P3), deterministic optimality (P4), constrained feasibility (P5), and nonstationarity (P6). * Part I (Necessity). I define six primitives X1–X6 as purely functional properties of decision rules: Objective Tracking, Cross-Context Safety Transfer, Global Attractor Exploration, Policy Simplification, Feasibility Projection, and Feedback Adaptation. For each, I construct an explicit environment in C and prove an unconditional Ω(T) regret lower bound for any decision-maker lacking that primitive.* Part II (Independence). For every ordered pair (i,j) with i≠j, I exhibit an explicit decision rule possessing Xj but lacking Xi that suffers Ω(T) regret on the matching environment. All thirty directed-pair results are shown to follow from one master theorem, verified on a single compound environment with full non-interference analysis.* Part III (Sequential Dependence). Necessity is domain-invariant — a structural failure is a structural failure no matter what "success" means to the decision-maker — which is why Parts I and II hold unconditionally. Sufficiency is not: what counts as success is supplied by the domain, not by the theorem, so a single closed-form sufficiency result covering every domain at once would have to either fix one arbitrary notion of success and stop being general, or say nothing of substance. Part III proves exactly what generalizes. I prove six Information Enhancement Theorems establishing that the six primitives compose into a directed information chain: possessing Xi strictly increases the mutual information available toward any goal variable at Xi+1's task. Each of the six links is established outright — a forward theorem, a reverse theorem, and a non-reversibility result — with the exact point where a domain's own definition of success enters the chain named explicitly, as an Implementation Obligation, rather than assumed away. The closing link, X6→X1, is grounded in an actual Doob martingale construction: given that the cycle-indexed posterior is a martingale, it converges almost surely to the truth across cycles — the precise sense in which the chain accumulates rather than resets. To this paper's knowledge, no prior formalization unifies this many independently-proven-necessary structural properties into a single machine-checked class with proven mutual independence across all of them. All mathematical work is provided in full transparency and independent verification is highly encouraged: the complete Lean formalization, with a passing build and every theorem cross-referenced to its exact identifier, is at github.com/M-Ismail-ZA/IsmailsPrimitives. For any feedback or collaboration, please contact me via the email address listed on the paper. Updated: 3 July 2026 (V6.1).
The rapid expansion of Internet of Things (IoT) applications has revealed limitations in traditional blockchain systems, particularly in scalability, energy efficiency, and computational overhead. IOTA, a Distributed Ledger Technology (DLT) based on Directed Acyclic Graph (DAG) structure known as the Tangle, offers a lightweight, scalable solution tailored for IoT environments. Key factor in IOTA’s performance and security is its tip selection mechanism, which manages transaction confirmation. This research introduces a novel approach combining Action Candidate-based Clipped Double Q-learning (AC-CDQ) with Biased Thompson Sampling (BTS) to improve tip selection. The proposed model addresses overestimation bias in Q-learning while enhancing robustness against adversarial attacks. Extensive simulations using a ledger simulator and the OpenAI Gym environment shows that model significantly reduces the average number of unapproved transactions. It consistently outperforms baseline methods, including Uniform Random Tip Selection (URTS), Markov Chain Monte Carlo (MCMC), and standard Q-learning, by confirming more transactions with improved reliability. Although the model incurs slightly higher computational costs, it delivers more accurate Q-value estimates and better learning stability. This work advances tip selection algorithms and supports the development of secure, efficient, and scalable distributed ledger systems. It demonstrates the potential of hybrid reinforcement learning techniques in future IoT-oriented ledger technologies.
Abstract: The Financial Technology (Fintech) sector is changing at a swift pace, as artificial intelligence (AI) is extending its influence. Greater complexity and global linkages are going to demand from fintech the power to rethink the integrity of its cybersecurity mechanisms and fraud tactics that have gotten intense up to a growing extent. The paper argues for the necessity of an "Algorithmic Fortress," an AI-driven cybernetic system incorporating all possible technologies targeted at securing digital financial networks against cyber-attacks and acts of financial fraud. The article delves into AI/ML, deep learning, anomaly detection through generative adversarial networks, etc., scope to predict battle, detect and fight problems. It does address adverse effects of AI risk, threatened system independence through synthetic identity fraud, application of AI for fraud detection in decentralized finance, DeFi, as well as the threat-hunting models that need to become autonomous. Supervised learning, unsupervised learning, and reinforcement learning are examination methodologies that are being applied in taking high recourse to the preservation of cybersecurity amongst their uncertainties. Our analysis will involve different experimentations of Python-based simulated attack scenarios to compare the two forms of cybersecurity. Also brought in are SmartArt visual representations revealed in multi-tier defensive architectures, combined with some strategic recommendations destined to protect future-facing fintech infrastructures from doing illicit deeds of algorithms. This study sketches possible solutions for securing the future-ready, trustworthy, and resilient fintech ecosystems once assisted by AI-enhanced, digital fortresses.
The emergence of AI agents and agentic systems represents a significant milestone in artificial intelligence, enabling autonomous systems to operate, learn, and collaborate in complex environments with minimal human intervention. This paper, drawing on multi-expert perspectives, examines the potential of AI agents and agentic systems to reshape industries by decentralizing decision-making, redefining organizational structures, and enhancing cross-functional collaboration. Specific applications include healthcare systems capable of creating adaptive treatment plans, supply chain agents that predict and address disruptions in real-time, and business process automation that reallocates tasks from humans to AI, improving efficiency and innovation. However, the integration of these systems raises critical challenges, including issues of attribution and shared accountability in decision-making, compatibility with legacy systems, and addressing biases in AI-driven processes. The paper concludes that while agentic systems hold immense promise, robust governance frameworks, cross-industry collaboration, and interdisciplinary research into ethical design are essential. Future research should explore adaptive workforce reskilling strategies, transparent accountability mechanisms, and energy-efficient deployment models to ensure ethical and scalable implementation.
The rapid adoption of artificial intelligence (AI) systems, such as predictive AI, generative AI, and explainable AI, is in contrast to the slower development and uptake of robotic AI systems. Dynamic environments, sensory processing, mechanical movements, power management, and safety are inherent complexities of robotic intelligence capabilities that can be addressed using novel AI approaches. The current AI landscape is dominated by machine learning techniques, specifically deep learning algorithms, that have been effective in addressing some of these challenges. However, these algorithms are subject to computationally complex processing and operational needs such as high data dependency. In this paper, we propose a computation-efficient and data-efficient framework for robotic motion intelligence (RMI) based on vector symbolic architectures (VSAs) and blockchain-based smart contracts. The capabilities of VSAs are leveraged for computationally efficient learning and noise suppression during perception, motion, movement, and decision-making tasks. As a distributed ledger technology, smart contracts address data dependency through a decentralized, distributed, and secure transactions ledger that satisfies contractual conditions. An empirical evaluation of the framework confirms its value and contribution towards addressing the practical challenges of robotic motion intelligence by significantly reducing the learnable parameters by 10 times while preserving sufficient accuracy compared to existing deep learning solutions.
Abstract When using autonomous robots for group foraging, it is extremely important to properly organize the exchange of information between members. In a decentralized system, it can be accomplished through pairwise interactions between closely located agents. The paper examines the role of obstacles and bottlenecks in creating an environment conducive to multiple exchanges. Reinforcement learning of groups with various exchange process organizations and obstacle distributions showed a weak impact on the results of randomly placed obstacles and a more significant effect of extended obstacles for both simulated and real robots. The role of data sharing turned out to be higher at the initial stage of the system’s operation and in changing operating environments. The results of the work can be used to organize data exchange when training groups of agents.
Distributed Control Multi-Agent Systems
Reinforcement Learning in Robotics
Advanced Research in Systems and Signal Processing
As more users seek generative AI (GAI) models to enhance work efficiency, GAI and Model-as-a-Service will drive transformative changes and upgrades across all industries. However, when users utilize GAI models provided by the service provider, they cannot be certain that the model’s quality matches the provider’s claims. Considering the need to protect intellectual property, the service provider will not disclose model details for user verification. To this end, we take the Internet of Vehicles as research background, proposing a zero knowledge model proof architecture based on UAVs. We also introduce a multiagent reinforcement learning algorithm to optimize the verification process. In specific, we first propose a verification scheme for the key operations of generative adversarial networks based on noninteractive zero knowledge proof. The zero knowledge proof architecture ensures that model parameters cannot be stolen during the verification process. After that, we propose an Age of Verification (AoV) metric to ensure the timeliness and freshness of zero knowledge proof. We also construct a tradeoff optimization problem between the energy consumption of UAV as a verifier and the AoV of edge servers as service providers, and transform the problem based on Lyapunov optimization theory. Following that, we propose an enhanced multiagent proximal policy optimization algorithm to enable the collaborative verification of edge servers by multiple UAVs. The algorithm simulation results demonstrate that the reward value of our proposed algorithm is over 10% higher than that of the standard algorithm, with a faster and more stable overall convergence speed. Additionally, the zero knowledge proof performance test results indicate that the verification delay in our proposed architecture is less than 500 ms during the verification phase, meeting practical requirements.
ABSTRACT This paper deals with the two‐player zero‐sum game problem, which is a bounded ‐gain robust control problem. Finding an analytical solution to the complex Hamilton‐Jacobi‐Issacs (HJI) equation is a challenging task. Hence, a novel Q‐learning algorithm for unknown continuous‐time (CT) affine‐in‐inputs nonlinear systems is proposed for generating an approximate solution to the HJI equation, which is valid in a local domain due to the use of a local approximator, that is, a Neural Network (NN) structure. The approach is model‐free and does not require the knowledge of system drift dynamics, and input and disturbance gains. The algorithm learns online from measurements of state variables in real time. To generate the local approximate solution of the HJI equation for the two‐player zero‐sum game problem for nonlinear systems, the proposed non‐iterative algorithm requires only a single critic NN instead of the commonly used triple NN approximator structure. A persistence of excitation condition is required to guarantee Uniformly Ultimately Boundedness (UUB) and convergence to the optimal solution. The effectiveness of the proposed Q‐learning approach for the two‐player zero‐sum game problem is demonstrated via simulations of a linear F‐16 aircraft plant and a highly complex nonlinear system. Proof of closed‐loop system stability is provided using Lyapunov Analysis, and convergence of the approximate solution to the true saddle‐point solution is guaranteed in a UUB‐sense.
Optimistic Rollup is a blockchain scalability solution. However, it assumes that transactions are correct unless challenged by a challenger, which leads to the concept of a challenge period. If there is no challenge during this time, the user needs to wait until the end of the challenge period, which is very inconvenient for the user. This is because challengers and task publishers are self-interested and aim to maximize their own benefits, leading to internal incentive issues within Optimistic Rollup. Although zero-knowledge proofs can effectively address this issue, ZK rollups face challenges such as high computational complexity and the need for transaction-specific proofs. Optimistic rollups have low computational requirements and simple design, making the design of an effective incentive mechanism a good choice. This paper proposes a reward-based incentive model to encourage rational workers to actively verify transactions based on game theory. The rewards in this incentive model are related to a time factor, where earlier verification yields greater rewards, motivating verifiers to verify transactions more quickly. We simulate under various proportions of false declarations and the results show that our method could incentivize the workers effectively to exert their best efforts.
Noor Al-Gburi, András Földvári, Kristóf Marussy, Oszkár Semeráth · 5 authors
Cross-organizational, blockchain-based distributed ledger networks in general, and those based on Hyperledger Fabric in particular, have an architecture which can be adapted to specific application requirements. However, network design can be a particularly challenging task, as the connection between architectural and deployment decisions and extra-functional properties can be subtle and the requirements may contradict each other, requiring trade-offs.
This chapter looks at the interaction of decentralized finance and artificial intelligence in the financial industry. The investigation begins with a review of the history and objectives, laying the groundwork for a thorough study. Section 2 goes into the foundations of DeFi, addressing key ideas and ecosystem components necessary for understanding how it works. Following that, Section 3 explores AI's revolutionary potential in finance, focusing on its role in financial services and the many AI technologies driving industry innovation. Section 4 focuses on real-world applications, emphasizing use cases such as automated trading algorithms, risk management, fraud detection, and decentralized oracle networks to demonstrate the practical ramifications of merging DeFi with AI. Section 5 then discusses the issues and considerations that come with this paradigm change, such as ethical and regulatory concerns, as well as security threats related with decentralized technology and AI-powered systems. Finally, Section 6 provides a forward-looking view by highlighting themes that will impact the future.
The objective of nature-inspired robotics is to augment the functionalities of robotic systems by imitating the intricate behavioral systems that are present in the natural world. By deriving inspiration from a wide range of ecosystems, this discipline applies biomechanical, ethological, and evolutionary principles to the development of robotics that exhibit efficient and adaptable actions. The concept of nature-inspired robotics originates from the observation that natural organisms navigate their environments with unparalleled adaptability, efficiency, and resilience. Researchers strive to emulate the complexities of biological systems in artificial agents through the process of decoding them, with the ultimate goal of improving the functionalities of robotic platforms. Nature-inspired robotics encompasses a wide range of domains, including soft robotics, swarm robotics, and bio-inspired algorithms. By incorporating knowledge from the study of ecosystem dynamics, animal behavior, and plant morphology, scientists intend to create robotic systems that can adapt autonomously to challenging and dynamic environments. The task involves effectively conveying the intricacies of natural behaviors through algorithmic frameworks and tangible manifestations. It is of utmost importance to surmount the intrinsic constraints of conventional robotics, including inflexible architectures and predetermined code, in order to develop robots capable of dynamically adapting to evolving circumstances, imitating collaborative intelligence, and seamlessly interacting with their environment. The principal aim is to design and construct robotic systems that replicate, and ideally exceed, the adaptability and efficiency of their biological counterparts. By amalgamating insights from various academic disciplines, the objective is to develop autonomous systems that can perform a wide array of duties, including environmental monitoring, search, and rescue, with minimal reliance on explicit human intervention. Nature-inspired robotic behavioral systems have exhibited encouraging progressions in the domains of self-organization, decentralized decision-making, and swarm intelligence. By imitating the behaviors of natural ecosystems, these robots demonstrate the capacity to revolutionize domains including precision agriculture, environmental exploration, and disaster response by adapting to unanticipated obstacles.
As Web3 projects leverage airdrops to incentivize participation, airdrop hunters tactically amass wallet addresses to capitalize on token giveaways. This poses challenges to the decentralization goal. Current detection approaches tailored for cryptocurrencies overlook non-fungible tokens (NFTs) nuances. We introduce ARTEMIS, an optimized graph neural network system for identifying airdrop hunters in NFT transactions. ARTEMIS captures NFT airdrop hunters through: (1) a multimodal module extracting visual and textual insights from NFT metadata using Transformer models; (2) a tailored node aggregation function chaining NFT transaction sequences, retaining behavioral insights; (3) engineered features based on market manipulation theories detecting anomalous trading. Evaluated on decentralized exchange Blur's data, ARTEMIS significantly outperforms baselines in pinpointing hunters. This pioneering computational solution for an emergent Web3 phenomenon has broad applicability for blockchain anomaly detection. The data and code for the paper are accessible at the following link: \hrefhttps://doi.org/10.5281/zenodo.10676801 doi.org/10.5281/zenodo.10676801.