Unreal Engine: Introduction to AI

After last week’s deeper dive into game effects, this week was AI fundamentals: AI Characters, AI Controllers, and the Navigation System (NavMesh) – the three pillars of making entities that can think, move, and collide.

Rather than re-invent the wheel, I took advantage of the existing codebases and examples from the course and uploaded them to the following repositories for future reference:

Notes below.

The Key Elements of AI

AI Character – the entity in the game world. Humanoids typically derive from the Character class (getting movement, health, collision); non-humanoids (drones, tanks) from Pawn
AI Controller – the “brain” for AI actors (move, attack). Distinct from the AI Character and exists separately – the same controller/pawn split the player side has had since the controllers week
Navigation System (NavMesh) – AI characters rely on the NavMesh to understand traversable space and calculate paths from A to B

What the split buys you

Delegation of tasks: the controller issues commands (move to this location), the character executes them
Separation of awareness: the character physically exists in the world; the controller handles high-level decision-making
Classic example: an AI guard patrols an area → the guard “sees” the player → the controller abandons patrolling and chases

Adding the NavMesh

Place a NavMeshBoundsVolume (Volumes section) to designate the nav mesh area - the volume acts as a container for the NavMesh
Paths are built automatically from the geometry inside the volume; green areas are walkable surfaces
Walls, stairs, and uneven terrain are all taken into account
View it by pressing P, or the Show Navigation console command

View modes and debugging

A RecastNavMesh actor is created automatically when the NavMesh is added – this does the actual generation work. Useful view modes on it:

Draw Polygon Costs - displays the cost agents pay for path calculations
Tile build time as heatmap - colour-codes tiles by how long they took to build
Heuristic Scale - tunes the pathfinding bias (see flag below)

Multiple agents of different sizes

Different-sized agents need multiple NavMeshes: small agents in tight spaces, larger agents where they fit
Adding an agent entry automatically creates an extra RecastNavMesh actor
Enable drawing on it to inspect; CycleNavDrawn cycles visibility between them at runtime
NavMesh assignment can be automatic – the pawn’s capsule size determines which NavMesh it uses

Custom Nav Volumes

Nav Modifier Volumes come with four pre-created setups:

Default - regular navigable area
Null - impassable
Obstacle – expensive to cross, only crossed when essential
Low Height – cannot be traversed

Specific regions can be assigned these properties to influence how AI behaves when crossing them.

For full control, inherit from NavArea to create a custom NavMesh modifier – altering the default crossing cost, the fixed area entry cost, and the display colour.

Nav Link Proxies

Areas that weren’t connected automatically can be linked manually:

NavLinkProxy – placed so its left/right points connect to the NavMesh; can constrain the direction of travel
Smart Nav Link Proxies – for complicated links, e.g. jumping up a ledge:
- Override ReceiveSmartLinkReached - triggers when an actor steps on the link’s start
- Set Smart Link is Relevant to true
- CalculateVelocity() with DestinationLocation and StartingLocation inputs, Velocity output
Auto-generation: tick Generate Nav Links on the RecastNavMesh actor, create a new actor with GeneratedNavLinksProxy as parent, set Link Proxy Class to it, and links that need jumping should generate automatically

Dynamic NavMeshes

NavMesh generation is Static by default – generated before runtime, never updated
Navigation Mesh → Runtime → Runtime Generation offers:
- Dynamic Modifiers Only – updates the NavMesh but only for Nav Modifiers
- Dynamic - adapts to everything including collisions; the most resource-intensive

Avoidance Systems

No Avoidance – agents don’t attempt to avoid each other; collisions happen
Reciprocal Velocity Obstacle (RVO) – each actor pushes away from other agents it encounters, adapting paths on the fly
Detour Crowd Manager – considers other actors when selecting paths; use a DetourCrowdAIController to control characters, and tune Project Settings → Crowd Manager (max agents, max agent radius)
MassCrowd / Mass AI - outperforms the others and scales to thousands of agents

Notes on things that stood out

The mental model that made this week click: the controller/pawn split for AI mirrors the player’s controller/pawn split exactly – meaning everything from the controllers and communication weeks transfers wholesale. An AI guard “seeing” a player is a perception problem bolted onto the same architecture you already know.

The avoid-system ladder (None → RVO → Detour Crowd → Mass) is really a scalability ladder – picking the right rung for your agent count matters more than any settings tweak within a rung.