Introduction
In classical AI planning, agents typically operate under the assumption of a fixed goal. However, real-world environments—particularly high-stakes maritime domains—are characterized by "exogenous events" such as sudden weather changes, partial observability of enemy forces, and continuous physical processes like fuel consumption. In these dynamic settings, a pre-defined plan can quickly become obsolete, and more critically, the original goal itself may become irrelevant or hazardous.
This research addresses the challenge of Goal-Driven Autonomy (GDA): the capacity for an agent to autonomously monitor its environment, explain discrepancies between expectations and observations, and dynamically prioritize or generate new goals. The authors introduce ARTUE (Autonomous Response to Unexpected Events), an agent designed to navigate these complexities within a high-fidelity Navy strategy simulation.
The ARTUE Framework
The core of ARTUE is a conceptual GDA framework that moves beyond simple replanning. While traditional online planning focuses on finding a new path to a static goal, GDA allows the agent to reason about the goal itself. The framework consists of four primary tasks:
- Discrepancy Detection: The agent compares real-time observations against the expectations generated by its planner.
- Explanation Generation: When a discrepancy is detected, the agent uses abductive reasoning—implemented via an Assumption-based Truth Maintenance System (ATMS)—to hypothesize the underlying cause (e.g., inferring the presence of a hidden enemy submarine).
- Goal Generation: Based on the inferred state, the agent applies domain-specific "principles" to formulate new goals.
- Goal Management: The agent prioritizes these goals based on their "intensity" or urgency to decide the next course of action.
ARTUE utilizes an extended version of the SHOP2 Hierarchical Task Network (HTN) planner, modified to support PDDL+. This allows the agent to reason about continuous time and effects, which is essential for modeling maritime movement and resource consumption.
Experimental Setup
The system was evaluated using the TAO (Tactical Action Officer) Sandbox, a high-fidelity simulator for naval operations. Three distinct scenarios were developed to test the agent's autonomy:
- Scouting: Identifying ships while managing the threat of a hidden submarine.
- Iceberg: Navigating cargo through a storm where an iceberg blocks a critical port, necessitating the rescue of other vessels.
- SubHunt: Tracking an enemy submarine that is actively deploying mines.
Results and Discussion
The authors conducted an ablation study comparing ARTUE against three baseline agents: PLAN1 (static execution), REPLAN (reactive replanning without explanation), and EXPLAIN (belief updating without goal change).
The results demonstrate that ARTUE significantly outperforms all baseline agents across all scenarios. For instance, in the SubHunt scenario, ARTUE achieved a performance score of 0.98, compared to 0.74 for the EXPLAIN agent. Statistical analysis confirmed these gains were significant ($p < .0001$).
A key insight from the study is that while explaining discrepancies improves performance (as seen in the EXPLAIN agent's superiority over REPLAN), the most substantial performance leap occurs when the agent can act on those explanations to autonomously generate and manage new goals.
Conclusion
ARTUE demonstrates the power of integrating explanation and goal reasoning in complex, unpredictable environments. By moving beyond "plan-and-execute" paradigms, GDA agents can handle surprises that would otherwise lead to mission failure. Future work will focus on reducing the "knowledge bottleneck" associated with manual goal-principle authoring and incorporating learning mechanisms so that agents can improve their explanatory models over time.
Summary updated