Navigation

Navigation with navcat

navcat builds a navmesh (the interconnected walkable polygons of your level) from raw geometry and answers pathfinding queries over it. Reach for it when something needs to move intelligently through the world — enemies routing around obstacles, NPCs patrolling, units ordered to a point, agents that must avoid walls and gaps. It is rendering-agnostic; navcat/three adapts it to Three.js.

Do you even need a navmesh? For "walk straight at the target on open, flat ground," a navmesh is overkill — steer directly. A navmesh earns its place once agents must route around obstacles, across uneven or multi-level terrain, or as a coordinated crowd. If your case is simpler, implementing your own steering is the right call.

Versions move — read the installed types

This skill routes you to the right capability and decision; it does not pin exact signatures, because navcat's API (export names, option fields, return shapes) shifts between versions. Before calling into it, read the installed package's exports/types (navcat, navcat/blocks, navcat/three) to confirm current names and fields. Treat the code below as shape, not contract.

The pipeline at a glance

1. Extract geometry — turn your walkable Three.js meshes (floor, terrain, obstacles) into positions/indices with getPositionsAndIndices(meshes) from navcat/three.
2. Generate the navmesh — one call: generateSoloNavMesh(input, options) from navcat/blocks for a single area (the default), or generateTiledNavMesh for large worlds and partial rebuilds.
3. Query — paths, nearest walkable point, random points, line-of-sight, over the generated navmesh.
4. Move agents — follow the path yourself each frame, or hand many agents to the crowd block.

import { findPath, DEFAULT_QUERY_FILTER } from 'navcat'
import { generateSoloNavMesh } from 'navcat/blocks'
import { getPositionsAndIndices, createNavMeshHelper } from 'navcat/three'

const [positions, indices] = getPositionsAndIndices([floor, ...obstacles])
const { navMesh } = generateSoloNavMesh({ positions, indices }, options) // see options below

scene.add(createNavMeshHelper(navMesh).object) // debug: see the walkable surface

const result = findPath(navMesh, start, end, halfExtents, DEFAULT_QUERY_FILTER)
const waypoints = result.path.map((p) => p.position) // [x,y,z][]

halfExtents is the search box used to snap start/end onto the mesh; if a query fails, your points are likely off the navmesh — widen it or snap first with the nearest-poly query.

Tuning generation

The generation options size the navmesh to your agent and trade detail against speed. You don't need to memorize the field list (read it from the installed SoloNavMeshOptions); you need the intuition for the few that matter:

• Agent size — walkable radius (clearance kept from walls), height (low ceilings excluded), climb (max step-up), and max walkable slope angle. Set these to your character.
• Resolution — cell size/height (voxel grid): smaller is finer but slower. A common start is cell size ≈ agent radius / 3.
• Simplification — region/edge/error limits control how blocky vs. smooth the polygons come out.

To customize an individual generation stage (tag areas, inject regions), navcat exposes the lower-level recast-style steps the one-call generators are built from — reach for those only when the generator can't express what you need.

Querying the navmesh

All queries take the navmesh and a query filter (DEFAULT_QUERY_FILTER unless you mark area costs/flags). The capabilities to route to:

• findPath — full path between two points as waypoints. findSmoothPath for corner-free paths along the surface. (For cached/dynamic movement, the node-path + straight-path pair gives finer control.)
• Nearest walkable point — snap a world position onto the mesh; the basis for "is this reachable" and for fixing failed queries.
• Random walkable point (anywhere, or within a radius of a center) — spawns, patrol destinations, wander targets.
• raycast along the surface — line-of-sight between two points on the mesh. Mind the no-hit convention: a clear ray reports its hit parameter as the package's "no wall" sentinel (a max-value, not 1.0 or 0) — confirm it in the installed type rather than assuming.
• moveAlongSurface — slide a position toward a target while staying on walkable area; a simple controller without full repathing.
• Off-mesh connections — explicit links across gaps the mesh can't (jumps, ladders, teleports).

Following a path each frame

findPath gives waypoints; moving the agent is your loop's job. Walk toward the current waypoint, advance when you reach it, re-path when the destination changes or the agent drifts off:

let waypoints = result.path.map((p) => p.position)
let i = 0

function updateAgent(pos, speed, delta) {
  if (i >= waypoints.length) return // arrived
  const t = waypoints[i]
  const dx = t[0] - pos[0], dz = t[2] - pos[2]
  const dist = Math.hypot(dx, dz)
  if (dist < 0.2) { i++; return }
  pos[0] += (dx / dist) * speed * delta
  pos[2] += (dz / dist) * speed * delta
}

Many agents → the crowd block

For groups (enemies swarming, units moving together), use the crowd API in navcat/blocks instead of hand-following paths — it does frame-distributed pathfinding, agent-to-agent and wall avoidance, and spacing for you. Keep a crowd from clumping into one overlapping blob:

• Give each agent a radius matching its body and let avoidance hold spacing. Do not hand-set every agent's velocity straight at the same point each frame — that overrides avoidance and stacks them.
• To swarm a target, give each agent a goal offset on a ring around it (spread by index/angle) so they surround it rather than fighting for its center.
• Tune avoidance strength and max acceleration so agents slide around each other instead of jittering or interpenetrating.

For advanced needs, copy the crowd block into your project and modify it.

Limitations — know when to fill the gap yourself

• The navmesh is static. It reflects the geometry at generation time. When walkable geometry changes (a door opens, terrain deforms), you must regenerate — generateSoloNavMesh rebuilds the whole area; tiled navmeshes rebuild only affected tiles. For a few moving obstacles, your own local steering/avoidance on top is often cheaper than constant rebuilds.
• Conventions: navcat follows OpenGL conventions (right-handed, counter-clockwise winding). If your data differs, transform coordinates in and out. Imported external navmeshes must share/index vertices between adjacent polygons.
• Keep collision and the navmesh aligned to the same shaped terrain — not a flat floor under sculpted ground (see the world skill), or agents will path through visible hills.
• Navmeshes are plain JSON — serialize a generated one to skip regeneration at load.