Multi-Agent Reinforcement Learning (MARL) has gained significant interest in recent years, enabling sequential decision-making across multiple agents in various domains. However, most existing explanation methods focus on centralized MARL, failing to address the uncertainty and nondeterminism inherent in decentralized settings. We propose methods to generate policy summarizations that capture task ordering and agent cooperation in decentralized MARL policies, along with query-based explanations for When, Why Not, and What types of user queries about specific agent behaviors. We evaluate our approach across four MARL domains and two decentralized MARL algorithms, demonstrating its generalizability and computational efficiency. User studies show that our summarizations and explanations significantly improve user question-answering performance and enhance subjective ratings on metrics such as understanding and satisfaction.

Autonomous vehicles (AVs) operate in complex environments where failures are consequential. Sophisticated machine learning models for perception and planning are key to overcoming at least part of that complexity, but their black-box nature complicates validation and verification (V V). The recent integration of Vision-Language-Action (VLA) models into AVs introduces a unique opportunity: besides generating trajectories, these models produce an explicit Chain-of-Thought (CoT) explaining their underlying rationale. This CoT provides a rich specification to cross-check model outputs and detect inconsistencies that may expose unsafe or unintended behavior. This paper assesses whether CoTs from a recent open driving VLA can support such monitoring. We curate DriveAlignBench, a specialized dataset from NVIDIA's Alpamayo 1.5 VLA for AVs containing 150 CoT-trajectory pairs, which we manually annotate for reliability, trajectory consistency, and safety. Our analysis reveals that 33.3

In-context reinforcement learning (ICRL) is an emerging RL paradigm where an agent, after pretraining, can adapt to out-of-distribution test tasks without any parameter updates, instead relying on an expanding context of interaction history. While ICRL has shown impressive generalization, safety during this adaptation process remains unexplored, limiting its applicability in real-world deployments where test-time behavior is expected to be safe. In this work, we propose SCARED: Safe Contextual Adaptive Reinforcement via Exact-penalty Dual, the first method that promotes safe adaptation of ICRL under the constrained Markov decision process framework. During the parameter-update-free adaptation process, our agent not only maximizes the reward but also keeps the accumulated cost within a user-specified safety budget. We also demonstrate that the agent actively reacts to the safety budget; with a higher safety budget, the agent behaves more aggressively, and with a lower safety budget the agent behaves more conservatively. Across challenging benchmarks, SCARED consistently enables safe and robust in-context adaptation, outperforming existing ICRL and safe meta-RL baselines.

Safe in-context reinforcement learning (ICRL) adapts online from interaction history without test-time parameter updates while controlling episode cost under a safety budget. Under out-of-distribution (OOD) deployment shifts, pretraining-only safe ICRL can give poor reward-safety tradeoffs because the remaining budget affects behavior only through frozen policy conditioning, not an explicit action-level check against predicted future cost. We propose a latent Q-Barrier shield that learns a context representation, latent dynamics, and an ensemble cost critic before deployment. Without parameter updates, the shield infers context from history and filters or softly reweights candidate actions using the remaining budget and predicted future cost. We prove a conditional, error-decomposed barrier-margin result: a Q-Barrier-satisfying action leaves the next latent-budget state with an approximately budget-safe continuation under the learned critic, up to Bellman and latent-prediction errors. Across five safe ICRL benchmarks, the shield improves deployment-time reward-safety tradeoffs over a strong safe-ICRL baseline: after a short context window, it achieves higher return in four of five benchmarks while matching or lowering average episode cost in all five.

This paper presents the outcomes of an open-floor session held at the 13th International Workshop on Engineering Multi-Agent Systems (EMAS 2025), aimed at co-developing a research roadmap for the EMAS community. Participants collaboratively identified and prioritised challenges in engineering large-scale, adaptive multiagent systems, particularly considering the need to engineer systems that can seamlessly integrate learning and reasoning. Through structured group discussions, four key challenges emerged: explainability in heterogeneous environments, environment modeling, handling dynamic contexts, and communication standardisation. For each of the challenges, participants proposed and ranked potential solutions based on impact and effort. The resulting roadmap highlights concrete research directions toward engineering intelligent, explainable, and interoperable multiagent systems that effectively integrate reasoning and learning in dynamic environments.

Safe Reinforcement Learning (RL) algorithms are typically evaluated under fixed training conditions. We investigate whether training-time safety guarantees transfer to deployment under distribution shift, using diabetes management as a safety-critical testbed. We benchmark safe RL algorithms on a unified clinical simulator and reveal a safety generalization gap: policies satisfying constraints during training frequently violate safety requirements on unseen patients. We demonstrate that test-time shielding, which filters unsafe actions using learned dynamics models, effectively restores safety across algorithms and patient populations. Across eight safe RL algorithms, three diabetes types, and three age groups, shielding achieves Time-in-Range gains of 13-14\% for strong baselines such as PPO-Lag and CPO while reducing clinical risk index and glucose variability. Our simulator and benchmark provide a platform for studying safety under distribution shift in safety-critical control domains.

Robotic manipulation is typically evaluated by task success, but successful completion does not guarantee safe execution. Many safety failures are temporal: a robot may touch a clean surface after contamination or release an object before it is fully inside an enclosure. We introduce SafeManip, a property-driven benchmark to explicitly evaluate temporal safety properties in robotic manipulation, moving beyond prior evaluations that largely focus on task completion or per-state constraint violations. SafeManip defines reusable safety templates over finite executions using Linear Temporal Logic over finite traces (LTLf). It maps observed rollouts to symbolic predicate traces and evaluates them with LTLf-based monitors. Its property suite covers eight manipulation safety categories: collision and contact safety, grasp stability, release stability, cross-contamination, action onset, mechanism recovery, object containment, and enclosure access. Templates can be instantiated with task-specific objects, fixtures, regions, or skills, allowing the same safety specifications to generalize across tasks and environments. We evaluate SafeManip on six vision-language-action policies, including π_0, π_0.5, GR00T, and their training variants, across 50 RoboCasa365 household tasks. Results show that even strong models often behave unsafely. Task-success gains do not reliably translate into safer execution: many successful rollouts remain unsafe, while longer-horizon or more complex tasks expose more violations. SafeManip provides a reusable evaluation layer for diagnosing temporal safety failures and measuring safe success beyond task completion.

As AI systems increasingly support human decision making, a central challenge is determining what information helps people recognize when to rely on AI predictions and when to question or override them. Across three controlled human-subject studies spanning abstract visual reasoning with RAVEN matrices and deductive logical reasoning with LSAT problems, we examine how different forms of AI support affect human–AI team performance. A multi-stage reveal study shows that AI predictions and explanations can affect objective accuracy and subjective confidence differently. In visual reasoning, LLM explanations do not improve accuracy beyond the predicted answer alone, and no additional support format significantly outperforms prediction-only support; predicted probabilities show the highest descriptive accuracy and error recovery, while a derived selective-automation policy provides a higher-performing reference benchmark. In language-based logical reasoning, by contrast, LLM explanations yield the highest accuracy and error recovery, outperforming expert-written explanations and probability-based support. These results show that no single support strategy is universally effective. Human–AI interfaces should instead be designed to support calibrated reliance and effective error recovery by matching the form of assistance to the task and the evidence available to users.

This paper addresses the challenges of computational accountability in autonomous systems, particularly in Autonomous Vehicles (AVs), where safety and efficiency often conflict. We begin by examining current approaches such as cost minimization, reward maximization, human-centered approaches, and ethical frameworks, noting their limitations addressing these challenges. Foreseeability is a central concept in tort law that limits the accountability and legal liability of an actor to a reasonable scope. Yet, current data-driven methods to determine foreseeability are rigid, ignore uncertainty, and depend on simulation data. In this work, we advocate for a new computational approach to establish foreseeability of autonomous systems based on the legal “BPL” formula. We provide open research challenges, using fully autonomous vehicles as a motivating example, and call for researchers to help autonomous systems make accountable decisions in safety-critical scenarios.

Reinforcement Learning (RL) has shown great promise in domains like healthcare and robotics but often struggles with adoption due to its lack of interpretability. Counterfactual explanations, which address "what if" scenarios, provide a promising avenue for understanding RL decisions but remain underexplored for continuous action spaces. We propose a novel approach for generating counterfactual explanations in continuous action RL by computing alternative action sequences that improve outcomes while minimizing deviations from the original sequence. Our approach leverages a distance metric for continuous actions and accounts for constraints such as adhering to predefined policies in specific states. Evaluations in two RL domains, Diabetes Control and Lunar Lander, demonstrate the effectiveness, efficiency, and generalization of our approach, enabling more interpretable and trustworthy RL applications.

Large Language Models (LLMs) are increasingly used for planning tasks, offering unique capabilities not found in classical planners such as generating explanations and iterative refinement. However, trust–a critical factor in the adoption of planning systems–remains underexplored in the context of LLM-based planning tasks. This study bridges this gap by comparing human trust in LLM-based planners with classical planners through a user study in a Planning Domain Definition Language (PDDL) domain. Combining subjective measures, such as trust questionnaires, with objective metrics like evaluation accuracy, our findings reveal that correctness is the primary driver of trust and performance. Explanations provided by the LLM improved evaluation accuracy but had limited impact on trust, while plan refinement showed potential for increasing trust without significantly enhancing evaluation accuracy.

There is a surge of interest in using formal languages such as Linear Temporal Logic (LTL) to precisely and succinctly specify complex tasks and derive reward functions for Reinforcement Learning (RL). However, existing methods often assign sparse rewards (e.g., giving a reward of 1 only if a task is completed and 0 otherwise). By providing feedback solely upon task completion, these methods fail to encourage successful subtask completion. This is particularly problematic in environments with inherent uncertainty, where task completion may be unreliable despite progress on intermediate goals. To address this limitation, we propose a suite of reward functions that incentivize an RL agent to complete a task specified by an LTL formula as much as possible, and develop an adaptive reward shaping approach that dynamically updates reward functions during the learning process. Experimental results on a range of benchmark RL environments demonstrate that the proposed approach generally outperforms baselines, achieving earlier convergence to a better policy with higher expected return and task completion rate.

Online planning for partially observable Markov decision processes (POMDPs) provides efficient techniques for robot decision-making under uncertainty. However, existing methods fall short of preventing safety violations in dynamic environments. This work presents a novel safe POMDP online planning approach that maximizes expected returns while providing probabilistic safety guarantees amidst environments populated by multiple dynamic agents. Our approach utilizes data-driven trajectory prediction models of dynamic agents and applies Adaptive Conformal Prediction (ACP) to quantify the uncertainties in these predictions. Leveraging the obtained ACP-based trajectory predictions, our approach constructs safety shields on-the-fly to prevent unsafe actions within POMDP online planning. Through experimental evaluation in various dynamic environments using real-world pedestrian trajectory data, the proposed approach has been shown to effectively maintain probabilistic safety guarantees while accommodating up to hundreds of dynamic agents.

There is a growing trend toward AI systems interacting with humans to revolutionize a range of application domains such as healthcare and transportation. However, unsafe human-machine interaction can lead to catastrophic failures. We propose a novel approach that predicts future states by accounting for the uncertainty of human interaction, monitors whether predictions satisfy or violate safety requirements, and adapts control actions based on the predictive monitoring results. Specifically, we develop a new quantitative predictive monitor based on Signal Temporal Logic with Uncertainty (STL-U) to compute a robustness degree interval, which indicates the extent to which a sequence of uncertain predictions satisfies or violates an STL-U requirement. We also develop a new loss function to guide the uncertainty calibration of Bayesian deep learning and a new adaptive control method, both of which leverage STL-U quantitative predictive monitoring results. We apply the proposed approach to two case studies: Type 1 Diabetes management and semi-autonomous driving. Experiments show that the proposed approach improves safety and effectiveness in both case studies.

Unseen shifts in environment dynamics, driven by hidden parameters such as friction or gravity, create a challenge for maintaining safety. We address this challenge by proposing Adaptive Shielding, a framework for safe reinforcement learning in constrained hidden-parameter Markov decision processes. A function encoder infers a low-dimensional representation of the underlying dynamics online from transition data, allowing the shield to adapt. To ensure safety during this process, we use a two-layer strategy. First, we introduce safety-regularized optimization that proactively trains the policy away from high-cost regions. Second, the adaptive shielding reactively uses the inferred dynamics to forecast safety risks and applies uncertainty-aware bounds using conformal prediction to filter unsafe actions. We prove that prediction errors in the shielding connect with bounds on the average cost rate. Empirically, across Safe-Gym benchmarks with varying hidden parameters, our approach outperforms baselines on the return-safety trade-off and generalizes reliably to unseen dynamics, while incurring only modest execution-time overhead. Code is available at https://github.com/safe-autonomy-lab/AdaptiveShieldingFE.

We present an optimization-based framework for robust permissive synthesis for Interval Markov Decision Processes (IMDPs), motivated by robotic decision-making under transition uncertainty. In many robotic systems, model inaccuracies and sensing noise lead to interval-valued transition probabilities. While robust IMDP synthesis typically yields a single policy and permissive synthesis assumes exact models, we show that robust permissive synthesis under interval uncertainty can be cast as a global mixed-integer linear program (MILP) that directly encodes robust Bellman constraints. The formulation maximizes a quantitative permissiveness metric (the number of enabled state-action pairs), while guaranteeing that every compliant strategy satisfies probabilistic reachability or expected reward specifications under all admissible transition realizations. To address the exponential complexity of vertex-based uncertainty representations, we derive a dualization-based encoding that eliminates explicit vertex enumeration and scales linearly with the number of successors. Experimental evaluation on four representative robotic benchmark domains demonstrates scalability to IMDPs with hundreds of thousands of states. The proposed framework provides a practical and general foundation for uncertainty-aware, flexibility-preserving controller synthesis in robotic systems.

In the evolving landscape of human-centered AI, fostering a synergistic relationship between humans and AI agents in decision-making processes stands as a paramount challenge. This work considers a problem setup where an intelligent agent comprising a neural network-based prediction component and a deep reinforcement learning component provides advice to a human decision-maker in complex repeated decision-making environments. Whether the human decision-maker would follow the agent's advice depends on their beliefs and trust in the agent and on their understanding of the advice itself. To this end, we developed an approach named ADESSE to generate explanations about the adviser agent to improve human trust and decision-making. Computational experiments on a range of environments with varying model sizes demonstrate the applicability and scalability of ADESSE. Furthermore, an interactive game-based user study shows that participants were significantly more satisfied, achieved a higher reward in the game, and took less time to select an action when presented with explanations generated by ADESSE. These findings illuminate the critical role of tailored, human-centered explanations in AI-assisted decision-making.

Partially observable Markov decision processes (POMDPs) have been widely used in many robotic applications for sequential decision-making under uncertainty. POMDP online planning algorithms such as Partially Observable Monte-Carlo Planning (POMCP) can solve very large POMDPs with the goal of maximizing the expected return. But the resulting policies cannot provide safety guarantees which are imperative for real-world safety-critical tasks (e.g., autonomous driving). In this work, we consider safety requirements represented as almost-sure reach-avoid specifications (i.e., the probability to reach a set of goal states is one and the probability to reach a set of unsafe states is zero). We compute shields that restrict unsafe actions which would violate the almost-sure reach-avoid specifications. We then integrate these shields into the POMCP algorithm for safe POMDP online planning. We propose four distinct shielding methods, differing in how the shields are computed and integrated, including factored variants designed to improve scalability. Experimental results on a set of benchmark domains demonstrate that the proposed shielding methods successfully guarantee safety (unlike the baseline POMCP without shielding) on large POMDPs, with negligible impact on the runtime for online planning.

Probabilistic model checking can provide formal guarantees on the behavior of stochastic models relating to a wide range of quantitative properties, such as runtime, energy consumption or cost. But decision making is typically with respect to the expected value of these quantities, which can mask important aspects of the full probability distribution such as the possibility of high-risk, low-probability events or multimodalities. We propose a distributional extension of probabilistic model checking, applicable to discrete-time Markov chains (DTMCs) and Markov decision processes (MDPs). We formulate distributional queries, which can reason about a variety of distributional measures, such as variance, value-at-risk or conditional value-at-risk, for the accumulation of reward until a co-safe linear temporal logic formula is satisfied. For DTMCs, we propose a method to compute the full distribution to an arbitrary level of precision, based on a graph analysis and forward analysis of the model. For MDPs, we approximate the optimal policy with respect to expected value or conditional value-at-risk using distributional value iteration. We implement our techniques and investigate their performance and scalability across a range of benchmark models. Experimental results demonstrate that our techniques can be successfully applied to check various distributional properties of large probabilistic models.

There is a growing interest in using formal languages such as Linear Temporal Logic (LTL) to specify complex tasks and reward functions for reinforcement learning (RL) precisely and succinctly. Nevertheless, existing methods often assign sparse rewards, which may require millions of exploratory episodes to converge to a quality policy. To address this limitation, we adopt the notion of task progression to measure the degree to which a task specified by a co-safe LTL formula is partially completed and design several reward functions to incentivize a RL agent to satisfy the task specification as much as possible. We also develop an adaptive reward shaping approach that dynamically updates reward functions during the learning process. Experimental results on a range of benchmark RL environments demonstrate that the proposed approach generally outperforms baselines, achieving earlier convergence to a policy with a higher success rate of task completion and a higher normalized expected discounted return.