HTN Planner AI decision making

Dawn Engine HTN Planner is a hierarchical task network (HTN) AI planning system: you author task-decomposition knowledge as a Domain, the planner solves a Plan against the agent’s WorldState, and HtnAgentComponent executes that plan with automatic replan on failure or environment change. Decisions are deterministic, explainable, and replayable.

Maturity: shipped. Provided by the HtnPlanner Gem (Code Gem; no hard Gem dependencies). After enablement, HtnSystemComponent mounts automatically; you still add HtnAgentComponent on AI entities and assign a compiled Domain asset.

It does not replace animation, navigation, or combat systems: operators call those capabilities. It also does not replicate WorldState over the network—authority usually owns the agent and runs planning locally.

What problem it solves

Game AI and simulation agents need goal-driven behavior that stays readable as tactics grow. Hand-rolled state machines and opaque ML policies become hard to explain (“why did it do that?”) and hard to author with designers. HTN Planner provides:

• A Domain graph of compound tasks, methods, and primitive operators;
• Deterministic planning (same Domain + same WorldState → byte-identical Plan);
• Per-entity plan execution with continuous mode, opportunistic replan, and failure backoff;
• Three extension channels (C++ operators, .htnnode presets/macros, script operators) without forking the Gem;
• HTN Canvas (standalone node editor), compile-time V1–V18 validation, ImGui live panel, and offline decomposition-trace replay.

With the Gem disabled, projects have no HTN planning path; O3DE itself ships no AI decision framework.

How it works

.htndomain (HTN Canvas graph)
 │
 ▼
AssetProcessor / HtnDomainBuilder ──► .htndomain_compiled (flat HtnDomainAsset)
 │
sensors / game code / scripts ──► WorldState │
 ▼
 HtnAgentComponent → FindPlan (DFS) → HtnPlanRunner
 │
 ▼
 IHtnOperator Begin/Tick → nav / anim / script / …

Two rules run through the Gem:

1. Model vs view: Domain data never depends on Qt/GraphCanvas; the editor graph is an authoring view; the builder flattens it to a read-only runtime asset shared by agents.
2. Extend via data/reflection: new tasks and operators register through C++, .htnnode JSON, or BehaviorContext—do not fork HtnPlanner for gameplay content.

Planning runs on a WorldState snapshot; effects are simulated during search, then committed at runtime by phase (PlanOnly / ApplyOnStart / ApplyOnSuccess). Decomposition is time-sliced across agents under a global per-frame step budget; results match one-shot FindPlan byte-for-byte.

In this section

Related capabilities

• Navigation3D HTN integration: optional Nav3D.* operators when both Gems are enabled; HtnPlanner itself has no hard dependency on it.
• Dawn Stream: remotely review levels that include HTN-driven agents.