R-HTN: Rebellious Online HTN Planning for Safety and Game AI

TL;DR

This work presents rebellious online HTN planning (R-HTN), enabling agents to reason with directives $\mathcal{D}$ and avoid unsafe states via $\mathcal{D}$-discrepancies $\delta(s)$. It introduces two agent variants, Nonadaptive and Adaptive, and an online planning algorithm that interleaves planning with execution and task repair when violations are anticipated or detected. Empirical results in O-RESCHU and MONSTER show Adaptive agents achieve more user-goals while incurring no state violations and lower penalties, compared to Compliant and Nonadaptive agents. The approach advances safe, personality-aware agent control, with implications for ethics, norms, and trustworthy human–agent interaction, and lays groundwork for richer directive representations and conflict resolution. $\mathcal{D}$-driven replanning and HTN-based task repair enable safe, goal-directed behavior under dynamic constraints, with potential extensions to numeric fluents, probabilistic outcomes, and directive hierarchies.

Abstract

We introduce online Hierarchical Task Network (HTN) agents whose behaviors are governed by a set of built-in directives \D. Like other agents that are capable of rebellion (i.e., {\it intelligent disobedience}), our agents will, under some conditions, not perform a user-assigned task and instead act in ways that do not meet a user's expectations. Our work combines three concepts: HTN planning, online planning, and the directives \D, which must be considered when performing user-assigned tasks. We investigate two agent variants: (1) a Nonadaptive agent that stops execution if it finds itself in violation of \D~ and (2) an Adaptive agent that, in the same situation, instead modifies its HTN plan to search for alternative ways to achieve its given task. We present R-HTN (for: Rebellious-HTN), a general algorithm for online HTN planning under directives \D. We evaluate R-HTN in two task domains where the agent must not violate some directives for safety reasons or as dictated by their personality traits. We found that R-HTN agents never violate directives, and aim to achieve the user-given goals if feasible though not necessarily as the user expected.

Figures (6)

• Figure 1: A randomly generated map where all five agents start from the location labeled "5s"
• Figure 2: Location of the five agents after four time ticks.
• Figure 3: Performance measures as a function of red zone spawn probability.
• Figure 4: The horizontal axis denotes the probability that the red zones spawn (with each tick). The vertical axis denotes the number of $\mathcal{D}$-discrepancies incurred, averaged over 100 episodes per agent type.
• Figure 5: Performance measures as a function of monster spawn probability.