Navigation and AI

Separate path data, movement execution, sensing, decision logic, spatial queries, interaction reservation, and local avoidance. Route only the current subsystem's details into context.

Select systems before building

• Use NavMesh for pathfinding over traversable collision. It does not decide goals or execute animations.
• Use Behavior Tree + Blackboard for event-driven, priority-ordered reactive decisions with observable aborts.
• Use StateTree for explicit hierarchical modes, transitions, sequences, and shared task data.
• Use a small component/enum/timer for trivial behavior; a graph asset is not mandatory.
• Use AI Perception for sensed stimuli and memory. Perception is evidence, not omniscient authoritative truth.
• Use EQS to generate, filter, and score environmental candidates. Do not run it continuously for facts a direct query already knows.
• Use Smart Objects to find and reserve interaction slots. The interactor executes the behavior and must release the claim.
• Use RVO or Detour Crowd, exclusively, for local avoidance. They complement pathfinding; they do not replace world paths.

Read system-selector-architecture.md first when choosing between these systems.

Route by task

Need	Load
Choose systems, ownership, update model	system-selector-architecture.md
Build NavMesh and tune agents/tiles/resolution	navmesh-generation-agents.md
Choose static/dynamic/invoker/World Partition navigation	dynamic-navigation-world-partition.md
Tune MoveTo, areas, filters, modifiers, links, failure recovery	movement-paths-links.md
Build Behavior Trees, Blackboards, decorators, services, tasks	behavior-trees.md
Build StateTree states, tasks, data, and transitions	statetree.md
Configure AI Perception and EQS	perception-eqs.md
Build reservable interactions and crowd avoidance	smart-objects-crowds.md
Debug, profile, network, and run test gates	debugging-performance.md
Apply common behavior patterns	patterns.md
Verify claims and maturity	sources.md

Execute in this order

1. State agent shapes, locomotion abilities, decision latency, maximum population, world/streaming model, network authority, and target hardware.
2. Establish correct collision and supported-agent settings; build the simplest static NavMesh that proves paths.
3. Choose navigation generation mode from actual world changes. Do not default to fully Dynamic.
4. Prove movement, arrival, path failure, off-mesh traversal, interruption, and recovery before decision graphs.
5. Choose one primary decision architecture and define its data contract.
6. Add Perception as event-fed observations; define forgetting and last-known-state semantics.
7. Add EQS only for candidate selection that needs scoring/filtering.
8. Add Smart Objects and crowd avoidance only after single-agent behavior is reliable.
9. Make authoritative AI decisions on the server in networked games; replicate observable gameplay results.
10. Debug selected agents and profile worst-case populations in a cooked target build.

Hard rules

• NavMesh green coverage does not prove an agent can move: controller, movement component, supported-agent selection, projection, path following, collision, links, and runtime availability must agree.
• Agent radius, height, slope, and step settings are gameplay contracts. Do not shrink them merely to force paths through invalid geometry.
• Do not make NavMesh fully Dynamic when static data, Dynamic Modifiers Only, streamed chunks, or invokers satisfy the requirement.
• World Partition NavMesh is explicitly Experimental in UE 5.8.
• Behavior Tree tasks must finish exactly once and clean up latent work when aborted. Use observer aborts deliberately.
• StateTree Tasks on active states run concurrently; the first completing task can trigger completion. Do not model an intended sequence as parallel tasks.
• Current UE 5.8 EQS documentation still describes the system as experimental. Bound query size/frequency and keep a fallback.
• AI Perception expiration is not the same as forgetting unless the project setting and state logic support it.
• Smart Objects contain opportunity/reservation data, not execution logic. Release claims on success, failure, abort, death, unload, or invalidation.
• RVO and Detour Crowd are independent alternatives and should not both control one agent.
• AI does not need to tick every decision every frame. Prefer events, observed data changes, bounded service intervals, and staggered work.

Answer contract

Return:

1. Selected systems and assumptions.
2. Authority and data ownership — server/client, controller/pawn, Blackboard/StateTree/context.
3. Navigation contract — agents, generation mode, areas/links, movement and failure behavior.
4. Exact actions and parameter effects.
5. Cancellation, unload, death, and invalidation recovery.
6. Performance/scaling consequences.
7. Verification — visible debug evidence and target-build pass criteria.

Do not invent universal NavMesh cell sizes, perception ranges, service intervals, or crowd limits. Derive them from agent geometry, level metrics, reaction design, population, and profiling.