Dynamic Mission Reconfiguration and Knowledge Injection for Autonomous Vehicles Decision-Making Frameworks During Field Operations

During offshore inspection missions, unexpected events are not rare. Each unplanned pause to diagnose or replan can be costly. Yet most autonomy stacks still respond by restarting the decision-making framework or falling back on small, offline precompiled contingency sets that cannot be extended during execution—forcing operators to trade flexibility for safety. Behavior Trees (BTs) remain the de facto method for structuring vehicle logic, but the need to list preconditions and enumerate scenarios leaves them vulnerable to the classic Qualification Problem: it is impossible to anticipate every context in which an action might fail. Data-driven extensions enable BTs to learn from experience, yet they overlook the richest source of domain expertise: the human operator. We introduce Coachable BT, a machine-coaching interface for runtime grafting: supervisors can splice new decision branches into the executing autonomy stack while the vehicle stays on task. Inspired by McCarthy’s concept of Elaboration Tolerance, Coachable BT treats operator advice as an additional tree rather than a patch to the failed node. Instead of relying only on a fixed library of preauthored contingencies, injected branches are stored in an auditable map as operator-authored runtime additions, and reused automatically when a similar failure recurs. The framework is validated in a hybrid autonomous surface vehicle–remotely operated vehicle inspection scenario simulated in Stonefish. Compared to other frameworks supporting runtime reconfiguration of decision-making logic, Coachable BTs enhance the user’s ability to seamlessly and transparently inject new knowledge during execution, while maintaining deterministic execution semantics. These results mark a step toward resilient, long-duration autonomous maritime missions that remain adaptable even under harsh field conditions.