Active Goal Recognition (AGR)

• AGR is a framework that infers an agent’s hidden goal by actively gathering evidence through cost-sensitive observations and interventions.
• It employs POMDP-based and reinforcement learning methods to balance observation costs with task progress in dynamic, uncertain environments.
• AGR leverages techniques such as environment design and adaptive sensing to reduce ambiguity and improve goal recognition accuracy.

Searching arXiv for recent and foundational papers on Active Goal Recognition. Active Goal Recognition (AGR) is the problem of inferring an agent’s hidden goal while the observer is itself an acting system whose choices affect what information becomes available. In contrast to passive goal recognition, AGR treats observation as a costly, decision-dependent process and embeds recognition into a broader control problem in which the observer may move, query, reconfigure the environment, or otherwise intervene to reduce uncertainty about the target’s goal (Amato et al., 2019, Zhang et al., 29 Jul 2025). Across the literature, AGR appears in several closely related forms: as a POMDP in which the observer plans information-gathering actions under uncertainty (Amato et al., 2019, Zhang et al., 29 Jul 2025, Saborio et al., 2024), as environment design that modifies a decision-making setting so that behavior becomes more revealing of goals (Kasumba et al., 2024), and as online adaptive recognition architectures that can serve as the recognition core of a larger active system (Elhadad et al., 14 May 2025, Nageris et al., 2024). A recurring theme is the separation between a goal-conditioned behavior model of the acting agent and a decision layer that exploits that model to choose informative actions.

1. Conceptual scope and formal problem variants

AGR extends classical goal recognition, where a goal recognition problem is given as a tuple such as $\langle \mathcal{D}, G, O \rangle$ or $T=(D,G,O)$, with $D$ a domain theory, $G$ a set of candidate goals, and $O$ an observation sequence, and the task is to identify the goal that best explains the observations (Elhadad et al., 14 May 2025, Nageris et al., 2024). In AGR, this passive inference problem is embedded in a setting in which the observer has its own state, action space, and costs, and must decide when to observe, where to position itself, or how to intervene (Amato et al., 2019, Zhang et al., 29 Jul 2025).

A canonical formulation models AGR as a POMDP. In “Active Goal Recognition” (Amato et al., 2019), the augmented state is $S=S_P\times S_T\times G_T$, where $S_P$ is the observer state, $S_T$ the target state, and $G_T$ the hidden target goal. The observer’s action space is partitioned into $A=A_P\cup A_O\cup A_D$, corresponding to task actions, observation actions, and decision or prediction actions. The reward function explicitly trades off observer task progress, observation costs, and prediction costs. This formulation makes recognition itself part of the observer’s planning objective rather than a standalone estimation routine (Amato et al., 2019).

A more explicitly probabilistic version appears in “Probabilistic Active Goal Recognition” (Zhang et al., 29 Jul 2025), which defines a multi-agent setting with an actor and an observer under the keyhole assumption. The observer maintains a joint belief

$T=(D,G,O)$0

over actor state and goal, with marginal goal belief

$T=(D,G,O)$1

The observer’s objective is to choose actions that sharpen this goal belief, using a belief-dependent reward

$T=(D,G,O)$2

which is maximized when the posterior over goals is concentrated (Zhang et al., 29 Jul 2025).

A broader formulation relevant to AGR is General Dynamic Goal Recognition (GDGR), introduced in “General Dynamic Goal Recognition” (Elhadad et al., 14 May 2025). GDGR models not one recognition problem but a sequence of goal recognition tasks over time, with changing domains, changing goal sets, and an information state $T=(D,G,O)$3 reused across tasks. Its objective,

$T=(D,G,O)$4

captures online, dynamic, adaptive recognition and is presented as a basis for AGR even though the observer in that work remains passive (Elhadad et al., 14 May 2025). This suggests that AGR can be viewed as GDGR plus an active decision layer.

2. Belief, observation, and information-seeking action

The central technical distinction between AGR and passive recognition is that observer actions change the observation process. In the probabilistic AGR framework of (Zhang et al., 29 Jul 2025), the observer controls its state $T=(D,G,O)$5, and observations are generated from an observation model $T=(D,G,O)$6. The observer therefore plans over how its future state will affect what it can see. The joint belief is updated by a predict–update recursion: $T=(D,G,O)$7 followed by

$T=(D,G,O)$8

This joint state–goal belief is more expressive than a goal-only posterior because it supports both positive evidence and negative evidence: not seeing the actor is itself informative when visibility depends on observer position (Zhang et al., 29 Jul 2025).

The same principle appears in AGR for robotic assistants. “Towards Intention Recognition for Robotic Assistants Through Online POMDP Planning” (Saborio et al., 2024) models a robot that must interleave support actions and information-gathering actions while assisting a human worker. The joint POMDP state is $T=(D,G,O)$9, where $D$0 is driven by a stochastic target process $D$1. Information-gathering actions such as “Perceive” or “Inspect” have explicit negative reward, while timely support actions can yield positive reward or prevent penalties incurred by the human’s failures. Recognition is thus not a separate estimation module but an emergent consequence of reward-maximizing planning under partial observability (Saborio et al., 2024).

A conceptually related but non-interventional formulation is “Goal-directed Planning and Goal Understanding by Active Inference” (Matsumoto et al., 2022). There, “goal understanding” is implemented in a variational recurrent neural network under active inference: given observed exteroceptive trajectories $D$2, the model minimizes a free-energy objective

$D$3

and outputs expected goals $D$4. Although the paper does not label this AGR, it unifies planning and recognition as two modes of inference in the same generative model, which is directly relevant to active settings where a recognizer must both understand and act (Matsumoto et al., 2022).

3. Planning-based AGR and POMDP control

The earliest explicit AGR formulation in the provided literature is “Active Goal Recognition” (Amato et al., 2019). Its experimental domains show how optimal behavior balances observation and task completion. In a corridor domain, the observer can either “work,” “observe,” or open a door believed to be the target’s destination. In a map domain, the observer chooses between movement, work, and help actions. AGR policies learned with SARSOP exploit the fact that early observations may have low information value and instead spend time on observer tasks until recognition becomes decision-relevant (Amato et al., 2019).

The probabilistic extension in (Zhang et al., 29 Jul 2025) replaces handcrafted reward engineering with a generic belief-based objective and solves the resulting problem online with Monte Carlo Tree Search. At decision nodes, AGR-MCTS uses UCB1-style action selection; at chance nodes, it samples next actor state, observer state, and observation, updates the joint belief $D$5, and backs up belief-based returns. This yields a domain-independent planner that performs comparably to a domain-specific greedy baseline in grid worlds, while the joint belief update itself substantially outperforms passive recognition on the same observation sequences (Zhang et al., 29 Jul 2025).

The robotic assistant formulation in (Saborio et al., 2024) uses online POMDP planning with POMCP and a relevance-based extension called RAGE. Partial Goal Satisfaction defines a potential

$D$6

with shaping term

$D$7

to mitigate delayed rewards during simulation. In maintenance and assembly scenarios, RAGE outperforms plain POMCP in average discounted return and variance, suggesting that reward shaping and relevance estimates help the observer time assistance and sensing more effectively (Saborio et al., 2024).

These planning-based AGR methods share three structural commitments. First, they model the observer as an explicit decision-maker. Second, they represent the target’s goal as hidden state. Third, they evaluate actions by their downstream impact on both information and task utility. A plausible implication is that AGR is best understood not as a special-purpose recognizer but as a belief-space control problem.

4. Recognition engines used within AGR

Although not all recent work implements active observers, several papers provide recognition engines that are explicitly positioned as suitable backbones for AGR.

“General Dynamic Goal Recognition” (Elhadad et al., 14 May 2025) proposes a generic algorithm with an information state $D$8, domain adaptation, goals adaptation, recognition inference, and information-state update. Its concrete implementation learns a goal-conditioned policy $D$9 once using SAC with HER and then, for each task, chooses between zero-shot transfer, few-shot fine-tuning, and cache reuse for each goal in the dynamic goal set. Recognition is then performed by selecting

$G$0

The paper states directly that this framework is well-suited as a backbone for AGR because it supports online observations, rapid adaptation, and reusable goal-conditioned behavior models (Elhadad et al., 14 May 2025).

“Goal Recognition using Actor-Critic Optimization” introduces DRACO, which learns one actor–critic pair per goal and derives distances between trajectories and goal policies, including the critic-based state-only distance

$G$1

the Wasserstein-style action distance

$G$2

and the Z-score distance

$G$3

Goal posteriors are then obtained by softmin-normalized distances and Bayes-style normalization. The paper explicitly notes that DRACO is a passive recognition oracle that could be inserted into an AGR loop and reevaluated online as new observations arrive (Nageris et al., 2024).

“Goal Recognition as Reinforcement Learning” (Amado et al., 2022) takes a related approach with one Q-function per goal. It defines a utility-based domain theory $G$4 and goal inference by

$G$5

with MaxUtil, KL-based, and divergence-point measures. The paper does not implement AGR, but the learned Q-functions and policies naturally serve as forward models of goal-conditioned behavior, which an active observer could exploit to select informative interventions (Amado et al., 2022).

“Leveraging Planning Landmarks for Hybrid Online Goal Recognition” (Wilken et al., 2023) offers a symbolic–statistical online recognizer combining planning landmarks with a Naive Bayes model. Its landmark completion score

$G$6

is updated incrementally as actions arrive, while a weighted combination with the Bayesian model yields a fast hybrid posterior-like score. This architecture is explicitly online and computationally efficient, making it a plausible inference module for AGR even though it remains passive (Wilken et al., 2023).

5. Environment design as active goal recognition

A distinct line of work treats AGR not as active sensing but as active modification of the environment to make goals easier to infer. “Data-Driven Goal Recognition Design for General Behavioral Agents” (Kasumba et al., 2024) formulates goal recognition design as: $G$7 where $G$8 is worst-case distinctiveness and $G$9 is a behavioral model. The metric is defined operationally as the number of initial actions that are identical for every pair of goals, i.e., the maximum length of a common action prefix across goal-conditioned paths (Kasumba et al., 2024).

This work is directly AGR-like because the observer or system designer intervenes in the environment before or during behavior to reduce ambiguity. Instead of selecting sensor actions, it selects layout modifications, blocked or unblocked cells, or object placements so that goal-conditional behaviors diverge earlier. The paper’s main technical contribution is to replace repeated exact $O$0 computation with a learned surrogate $O$1, enabling gradient-guided discrete optimization over environments (Kasumba et al., 2024).

Its broader significance for AGR is twofold. First, it weakens the usual optimal-agent assumption by allowing hyperbolically discounted and data-driven human behavior models. Second, it grounds the value of environment design with human-subject experiments: environments optimized using a learned human behavior model yield lower overlapping path prefixes and higher Bayesian inference accuracy on actual human trajectories (Kasumba et al., 2024). This suggests that AGR interventions can target not only what the observer sees, but also the geometry of the decision problem itself.

A related implication emerges from landmark-based work such as (Pereira et al., 2019) and (Gusmão et al., 2021), although these papers are not themselves AGR algorithms. Planning landmarks identify necessary subgoals, and landmark uniqueness measures how discriminative a landmark is across candidate goals. In an active setting, these structures can guide sensing or design choices toward observations that test high-uniqueness landmarks first. This suggests a bridge between symbolic goal structure and information value.

6. Learning-based, probabilistic, and human-centered directions

AGR increasingly intersects with data-driven recognition and human behavior modeling. “Imagination-Augmented Deep Learning for Goal Recognition” (Duhamel et al., 2020) augments recurrent neural networks with planner-derived cost gradients

$O$2

or heuristic deviations

$O$3

to improve recognition from partial trajectories. The planner acts as an imagination module, producing cost-derived features that summarize possible futures. While active control is absent, such goal posteriors are natural belief states for AGR policies that need fast updates under noisy and suboptimal behavior (Duhamel et al., 2020).

“Human Goal Recognition as Bayesian Inference: Investigating the Impact of Actions, Timing, and Goal Solvability” (Zhang et al., 2024) expands the evidential basis of recognition beyond actions. It models a goal recognition problem with timing information as $O$4, where observations contain action–time pairs $O$5, and uses

$O$6

Its key empirical finding is that actions dominate human inference, but timing and solvability matter when actions are uninformative (Zhang et al., 2024). For AGR, this supports a modality-selection view of active sensing: action information should be exploited first, while timing or solvability cues should be sought when action evidence is weak.

“Data-Driven Goal Recognition in Transhumeral Prostheses Using Process Mining Techniques” (Su et al., 2023) offers an assistive-control perspective closely aligned with AGR. Continuous sensor streams are transformed into discrete events, goal-specific Petri nets are discovered, and alignment-based weights

$O$7

are converted into goal probabilities

$O$8

Although the work does not implement a full active controller, it stresses that being less confident when wrong is beneficial because the prosthesis should not commit too aggressively to an incorrect goal. This is directly relevant to AGR systems whose actions depend on confidence calibration (Su et al., 2023).

The active inference formulation in (Matsumoto et al., 2022) similarly points toward a human-like AGR architecture in which the same generative model supports both action generation and goal understanding. A plausible implication is that AGR benefits from architectures where recognition and intervention share latent dynamics rather than being engineered as separate modules.

7. Empirical themes, limitations, and open problems

Several empirical regularities recur across the literature. First, negative evidence matters. Joint belief update in (Zhang et al., 29 Jul 2025) outperforms passive methods largely because non-detections are informative when the observer controls visibility. Second, fast reusable behavior models matter. Goal-conditioned RL in (Elhadad et al., 14 May 2025) and actor–critic recognition in (Nageris et al., 2024) shift computation offline and support cheap online reevaluation. Third, purely optimal planning assumptions are often brittle. Data-driven environment design explicitly targets general behavioral agents (Kasumba et al., 2024), and Bayesian analyses of human inference show that timing and solvability affect judgments in ways not captured by classical optimality-based models (Zhang et al., 2024).

Common limitations are equally clear. Many AGR systems assume known dynamics, fixed candidate goal sets, or full observability of relevant variables (Elhadad et al., 14 May 2025, Zhang et al., 29 Jul 2025). Planning-based POMDP formulations are computationally demanding and often demonstrated only in modest domains (Amato et al., 2019, Saborio et al., 2024). RL-based recognizers may require a simulator and typically scale linearly with the number of goals when one policy is trained per goal (Nageris et al., 2024). Environment design methods depend on the fidelity of the learned behavior model and on training-distribution coverage (Kasumba et al., 2024).

A recurring open problem is the integration of recognition with intervention under richer uncertainty. The papers collectively point toward several directions. One is multi-domain, continual, or lifelong AGR, anticipated by GDGR’s information-state formulation (Elhadad et al., 14 May 2025). Another is richer probabilistic belief-space planning that handles partial observability, unknown dynamics, or multi-agent interactions while retaining tractable online control (Zhang et al., 29 Jul 2025, Saborio et al., 2024). A third is better alignment with human behavior: bounded rationality, variable timing, and unsolvable goals are already shown to affect inference (Zhang et al., 2024), and environment design demonstrates that tailoring interventions to actual human behavior can materially improve recognizability (Kasumba et al., 2024).

Taken together, the literature defines AGR as the convergence of online goal inference, active information gathering, and decision-theoretic control. The field now contains explicit probabilistic POMDP formulations (Amato et al., 2019, Zhang et al., 29 Jul 2025), active environment-design approaches (Kasumba et al., 2024), scalable learning-based recognition cores (Elhadad et al., 14 May 2025, Nageris et al., 2024, Amado et al., 2022), and human-centered generative and Bayesian models (Matsumoto et al., 2022, Zhang et al., 2024). This suggests that AGR is less a single algorithmic technique than a family of architectures organized around one principle: the observer should act so as to make goals easier to infer, and should infer goals so as to act better.