AI, Pathfinding & NPC Behavior
Recast & Detour, A* Pathfinding, behavior trees, ML-Agents, Convai, Inworld.
Navigation meshes, pathfinding, behavior trees, GOAP / utility AI, and the new wave of LLM-driven NPCs. Pairs with 3D engines, code-first engines, and AI / LLM for the chat-NPC angle. For full landscape see Game Development.
Navigation meshes / pathfinding
- ★ Recast & Detour — free OSS, the open NavMesh standard. Originally by Mikko Mononen; the algorithm under Unity's NavMesh, Unreal's Recast wrapper, Godot's built-in nav. Pure C++.
- A Pathfinding Project* (Aron Granberg) — free or paid Pro for Unity. The Unity community standard above the built-in NavMesh.
- NavMeshPlus — free OSS Unity 2D NavMesh extension.
- Unity NavMesh + NavMesh Components (now official package) — built in, Recast under the hood.
- Unreal NavMesh — built in, Recast wrapper.
- Godot NavigationServer3D / NavigationAgent — built in (Godot 4 navigation overhauled).
- Pathfinding.js — free OSS JS A*/Dijkstra/etc.
Behavior systems
- Behavior Trees — built into Unreal (BT) and many Unity assets.
- Behavior Designer (Unity) — paid, popular Unity BT.
- AI Tree (Unity) — paid, modern BT/utility AI.
- NodeCanvas (Unity) — paid BT/state machine/dialogue.
- Behaviac — free OSS BT, cross-engine.
- Panda BT — free OSS Unity BT.
- GOAP libraries — Goal-Oriented Action Planning crates / packages for many engines.
- Utility AI — Apex Utility AI (paid Unity); GOAP and Utility ideas often hand-rolled.
State machines
- See state machines for general; in games, NodeCanvas, Animator State Machine (Unity), and Unreal's Gameplay Ability System (GAS) cover it.
Machine learning
- ★ Unity ML-Agents — free OSS, official Unity. Train RL agents in Unity environments with PyTorch under the hood.
- MLAgents-Cloud — paid scaling.
- OpenAI Gym + custom env — research-flavored.
- Godot RL Agents — free OSS Godot ML-Agents equivalent.
LLM-driven NPCs (2024+ wave)
- ★ Inworld AI — paid, real-time character LLM with goals/personality/memory. Unity + Unreal SDK. Used in demos far more than shipped games but improving.
- Convai — paid, similar to Inworld, more focus on knowledge bases.
- NVIDIA ACE — paid platform combining LLM + Audio2Face + Riva voice.
- Ollama / LM Studio + custom integration — free, local LLMs, run NPC dialogue offline.
- OpenAI / Anthropic / Gemini APIs — paid, but flexible. License + privacy considerations matter for games.
The honest 2026 take: LLM NPCs are shipping in tech demos, modding scenes, and a handful of indie titles. They are not yet the default; latency, cost, hallucination, and game-design fit (you usually want NPCs to obey the script) are still problems. Expect this to flip within 2 years.
Sensing / perception
- Unity Perception package — free, generates synthetic data for ML.
- AI Sensors — paid Unity asset, vision/hearing for AI.
Crowd / flow
- Recast / Detour Crowd — built into Recast.
- DetourCrowd in Unreal/Unity wrappers.
- Massive (Wētā) — paid AAA crowd sim.
Pricing summary
- Free / OSS: Recast & Detour, NavMeshPlus, A* Free, Behaviac, Panda BT, ML-Agents, Pathfinding.js.
- Paid Unity assets: A* Pro, Behavior Designer, NodeCanvas, AI Tree.
- Paid SaaS: Inworld, Convai, NVIDIA ACE, OpenAI/Anthropic.
Pick this if…
- Default NavMesh (Unity/Unreal/Godot): built-in (which is Recast under the hood).
- Unity, complex pathfinding (grids, hierarchical): A* Pathfinding Project.
- Behavior trees, Unity, paid: Behavior Designer.
- Behavior trees, free: Unreal BT (built in) or Behaviac.
- Game-RL training: ML-Agents (Unity) or Godot RL Agents.
- Experimental LLM NPCs: Inworld or Convai for hosted; Ollama for local.
- 2D pathfinding: A* Free or NavMeshPlus.