Learning per-domain generalizing policies is a key challenge in learning for planning. Standard approaches learn state-value functions represented as graph neural networks using supervised learning on optimal plans generated by a teacher planner. In this work, we advocate for learning Q-value functions instead. Such policies are drastically cheaper to evaluate for a given state, as they need to process only the current state rather than every successor. Surprisingly, vanilla supervised learning of Q-values performs poorly as it does not learn to distinguish between the actions taken and those not taken by the teacher. We address this by using regularization terms that enforce this distinction, resulting in Q-value policies that consistently outperform state-value policies across a range of 10 domains and are competitive with the planner LAMA-first.

Interfaces for human oversight must effectively support users' situation awareness under time-critical conditions. We explore reinforcement learning (RL)-based UI adaptation to personalize alerting strategies that balance the benefits of highlighting critical events against the cognitive costs of interruptions. To enable learning without real-world deployment, we integrate models of users' gaze behavior to simulate attentional dynamics during monitoring. Using a delivery-drone oversight scenario, we present initial results suggesting that RL-based highlighting can outperform static, rule-based approaches and discuss challenges of intelligent oversight support.

Neural networks are increasingly important to learn action policies. Policy predicate abstraction (PPA) verifies safety of such a neural policy pi by over-approximating the state space subgraph induced by pi and using counterexample-guided abstraction refinement (CEGAR) to iteratively refine the abstraction. So far, PPA verifies safety in non-deterministic systems. This work extends PPA to probabilistic verification. Extending the abstract state space computation is relatively straightforward. Abstraction refinement, however, becomes substantially more complex, due to the more intricate form of counterexamples and the various sources of spuriousness it entails. We tackle this challenge by drawing inspiration from prior work on probabilistic CEGAR, empowering it to deal with neural pi. The resulting algorithm decides whether pi is safe with respect to a desired upper bound on unsafety probability. Invoking the algorithm incrementally, we can also derive upper and lower bounds automatically. Our experiments show that these algorithms can derive non-trivial bounds, whereas encodings into state-of-the-art probabilistic model checkers turn out to be ineffective.

PDDL, the de-facto standard language in the AI Planning community, is designed to specify planning domains: sets of instances that share the same predicates and action schemas. Yet it does not provide any means to specify the actual instance set, i.e., legality constraints on initial states and goal conditions, as well as possibly domain subset constraints specifying an instance subset we are interested in. One consequence of this is that instance generation has always been ad-hoc, with manually written domain- and subset-specific instance generators. Recent work has started to address this, through reasoning and learning methods that however suffer from scalability limitations. Here we introduce an alternative approach, leveraging LLMs to generate instance-generation programs, with built-in soundness guarantees through prescribed checks. We show that these automatically generated instance generators return large numbers of sound and diverse instances efficiently.

Learning per-domain generalizing policies is a key challenge in learning for planning. Standard approaches learn state-value functions represented as graph neural networks using supervised learning on optimal plans generated by a teacher planner. In this work, we advocate for learning Q-value functions instead. Such policies are drastically cheaper to evaluate for a given state, as they need to process only the current state rather than every successor. Surprisingly, vanilla supervised learning of Q-values performs poorly as it does not learn to distinguish between the actions taken and those not taken by the teacher. We address this by using regularization terms that enforce this distinction, resulting in Q-value policies that consistently outperform state-value policies across a range of 10 domains and are competitive with the planner LAMA-first.

LLMs have recently been used to generate Python programs representing generalized plans in PDDL planning, i.e., plans that generalize across the tasks of a given PDDL domain. Previous work proposed a framework consisting of three steps: the LLM first generates a summary and then a strategy for the domain, both in natural language, and then implements that strategy as a Python program, that gets debugged on example planning tasks. In that work, only one strategy is generated and passed directly to the program generation. If the strategy is incorrect, its implementation will therefore result in an incorrect generalized plan. Here, we introduce an approach that generates the strategy in the form of pseudocode and enables automatic debugging of the pseudocode, hence allowing us to identify and fix errors prior to the generation of the generalized plan itself. Additionally, we extend the Python debugging phase with a reflection step prompting the LLM to pinpoint the reason for the observed plan failure. Finally, we take inspiration from LLM code generation to produce several program variants and pick the best one. Running experiments on 17 benchmark domains with two reasoning and two non-reasoning LLMs, we show that these extensions substantially improve the quality of the generalized plans. Our best performing configuration achieves an average coverage of 82% across the domains.

Heuristic search is a prominent method for plan generation in classical planning. Here we address its use for a new problem that we baptize action optimality checking (AOC): checking whether a given action a is optimal in a given state s. AOC has various potential uses, e.g. quality assurance for learned action policies through checking example policy decisions. A vanilla algorithm for AOC is to run two A(star) searches, on each of s and the outcome state s of applying a. We show that one can do much better than this. We introduce early termination criteria across multiple searches. Beyond this, we introduce AOCA(star), which performs a single search on s that gives preference to paths going through s. Our experiments show that AOCA(star) is superior to the vanilla algorithm as well as other multiple search configurations, consistently across three different state-of-theart heuristic functions.

Reinforcement learning is becoming ever more prominent in solving combinatorial search problems, in particular ones where states are images. Prior work has devised action-policy testing methodology, that identifies so-called bug states where policy performance is sub-optimal. Here we show how to leverage this methodology during the RL process, using action-policy testing to find bugs and injecting those as alternate start states for the training runs. Running experiments across six 2D games, we find that our testing-guided training often achieves similar expected reward while reducing the number of bugs.

Testing is a natural approach to assess the quality of learned action policies p. Prior work introduced policy testing in AI planning as searching for bugs in p, that is, states where p is sub-optimal with respect to a given testing objective. Beyond quality assurance, an obvious application of these methods is policy selection: given several p to choose from, we can use testing to select the "least buggy" one. Here, we integrate testing-based policy selection into the training process. This includes making more informed decisions when selecting the final policy after training, as well as choosing more promising intermediate policies during the training process. Our experiments with ASNets action policies show that integrating testing allows us to more reliably obtain good-quality policies.

Large language models (LLMs) have revolutionized a large variety of NLP tasks. An active debate is to what extent they can do reasoning and planning. Prior work has assessed the latter in the specific context of PDDL planning, based on manually converting three PDDL domains into natural language (NL) prompts. Here we automate this conversion step, showing how to leverage an LLM to automatically generate NL prompts from PDDL input. Our automatically generated NL prompts result in similar LLM-planning performance as the previous manually generated ones. Beyond this, the automation enables us to run much larger experiments, providing for the first time a broad evaluation of LLM planning performance in PDDL. Our NL prompts yield better performance than PDDL prompts and simple template-based NL prompts. Compared to symbolic planners, LLM planning lags far behind; but in some domains, our best LLM configuration scales up further than A* using LM-cut.

Recent work has shown that successful per-domain generalizing action policies can be learned. Scaling behavior, from small training instances to large test instances, is the key objective; and the use of validation instances larger than training instances is one key to achieve it. Prior work has used fixed validation sets. Here, we introduce a method generating the validation set dynamically, on the fly, increasing instance size so long as informative and feasible.We also introduce refined methodology for evaluating scaling behavior, generating test instances systematically to guarantee a given confidence in coverage performance for each instance size. In experiments, dynamic validation improves scaling behavior of GNN policies in all 9 domains used.

Recent work has introduced methodology for testing learned action policies in AI Planning, aiming to effectively identify bug states where policy behavior is suboptimal. While this work focused on cost-optimality in classical planning, here we apply the core ideas to safety testing in planning with initial-state and action-outcome non-determinism. We cover the entire testing pipeline, introducing fuzzing algorithms to find unsafe policy runs, as well as test oracles to identify bugs where such unsafe behavior could be avoided. Going beyond the previous framework, we introduce a final step to the pipeline, identifying faults which we define to be specific policy decisions - state/action pairs transitioning from a safe state (where a safe policy exists) to an unsafe state (where no such policy exists). We adapt a range of known algorithms for these purposes, including also approximate ones bounding the number of times we are allowed to diverge from the learned policy. We run comprehensive experiments evaluating each part of our pipeline. Key takeaways are that safety testing can be quite cheap, in contrast to cost-optimality testing; and that variants of Tarjan's algorithm tend to be highly effective for this purpose.

Adaptive agent design offers a way to improve human-AI collaboration on time-sensitive tasks in rapidly changing environments. In such cases, to ensure the human maintains an accurate understanding of critical task elements, an assistive agent must not only identify the highest priority information but also estimate how and when this information can be communicated most effectively, given that human attention represents a zero-sum cognitive resource where focus on one message diminishes awareness of other or upcoming information. We introduce a theoretical framework for adaptive signalling which meets these challenges by using principles of rational communication, formalised as Bayesian reference resolution using the Rational Speech Act (RSA) modelling framework, to plan a sequence of messages which optimise timely alignment between user belief and a dynamic environment. The agent adapts message specificity and timing to the particulars of a user and scenario based on projections of how prior-guided interpretation of messages will influence attention to the interface and subsequent belief update, across several timesteps out to a fixed horizon. In a comparison to baseline methods, we show that this effectiveness depends crucially on combining multi-step planning with a realistic model of user awareness. As the first application of RSA for communication in a dynamic environment, and for human-AI interaction in general, we establish theoretical foundations for pragmatic communication in human-agent teams, highlighting how insights from cognitive science can be capitalised to inform the design of assistive agents.

This paper presents the demonstrator developed in the TUPLES European Union research project for assisting human operators at Airbus to plan Beluga cargo ground logistic operations. The demonstrator features techniques providing robust, explainable, and safe decisions, which all contribute to making our decision-support system trusted by the operators. We have also worked on various planning methods to scale up to the size of the real industrial problem, including hybrid machine learning and symbolic algorithms. We demonstrate the software that was tested by Airbus operators during a user study in Finkenwerder’s production site in May 2025.

In many automated planning applications, action costs can be hard to specify. An example is the time needed to travel through a certain road segment, which depends on many factors, such as the current weather conditions. A natural way to address this issue is to learn to predict these parameters based on input features (e.g., weather forecasts) and use the predicted action costs in automated planning afterward. Decision-Focused Learning (DFL) has been successful in learning to predict the parameters of combinatorial optimization problems in a way that optimizes solution quality rather than prediction quality. This approach yields better results than treating prediction and optimization as separate tasks. In this paper, we investigate for the first time the challenges of implementing DFL for automated planning in order to learn to predict the action costs. There are two main challenges to overcome: (1) planning systems are called during gradient descent learning, to solve planning problems with negative action costs, which are not supported in planning. We propose novel methods for gradient computation to avoid this issue. (2) DFL requires repeated planner calls during training, which can limit the scalability of the method. We experiment with different methods approximating the optimal plan as well as an easy-to-implement caching mechanism to speed up the learning process. As the first work that addresses DFL for automated planning, we demonstrate that the proposed gradient computation consistently yields significantly better plans than predictions aimed at minimizing prediction error; and that caching can temper the computation requirements.

Neural networks (NN) are an increasingly important representation of action policies pi. Applicability filtering is a commonly used practice in this context, restricting the action selection in pi to only applicable actions. Policy predicate abstraction (PPA) has recently been introduced to verify safety of neural pi, through over-approximating the state space subgraph induced by pi. Thus far however, PPA does not permit applicability filtering, which is challenging due to the additional constraints that need to be taken into account. Here we overcome that limitation, through a range of algorithmic enhancements. In our experiments, our enhancements achieve several orders of magnitude speed-up over a baseline implementation, bringing PPA with applicability filtering close to the performance of PPA without such filtering.

Heuristic search guides the exploration of states via heuristic functions h estimating remaining cost. Symbolic search instead replaces the exploration of individual states with that of state sets, compactly represented using binary decision diagrams (BDDs). In cost-optimal planning, heuristic explicit search performs best overall, but symbolic search performs best in many individual domains, so both approaches together constitute the state of the art. Yet combinations of the two have so far not been an unqualified success, because (i) h must be applicable to sets of states rather than individual ones, and (ii) the different state partitioning induced by h may be detrimental for BDD size. Many competitive heuristic functions in planning do not qualify for (i), and it has been shown that even extremely informed heuristics can deteriorate search performance due to (ii).Here we show how to achieve (i) for a state-of-the-art family of heuristic functions, namely potential heuristics. These assign a fixed potential value to each state-variable/value pair, ensuring by LP constraints that the sum over these values, for any state, yields an admissible and consistent heuristic function. Our key observation is that we can express potential heuristics through fixed potential values for operators instead, capturing the change of heuristic value induced by each operator. These reformulated heuristics satisfy (i) because we can express the heuristic value change as part of the BDD transition relation in symbolic search steps. We run exhaustive experiments on IPC benchmarks, evaluating several different instantiations of potential heuristics in forward, backward, and bi-directional symbolic search. Our operator-potential heuristics turn out to be highly beneficial, in particular they hardly ever suffer from (ii). Our best configurations soundly beat previous optimal symbolic planning algorithms, bringing them on par with the state of the art in optimal heuristic explicit search planning in overall performance.

Testing was recently proposed as a method to gain trust in learned action policies in classical planning. Test cases in this setting are states generated by a fuzzing process that performs random walks from the initial state. A fuzzing bias attempts to bias these random walks towards policy bugs, that is, states where the policy performs sub-optimally. Prior work explored a simple fuzzing bias based on policy-trace cost. Here, we investigate this topic more deeply. We introduce three new fuzzing biases based on analyses of policy-trace shape, estimating whether a trace is close to looping back on itself, whether it contains detours, and whether its goal-distance surface does not smoothly decline. Our experiments with two kinds of neural action policies show that these new biases improve bug-finding capabilities in many cases.