## Generating 2D Navmeshes

Peer Reviewed by jbadams, Dave Hunt, Josh Vega

navmesh pathfinding ai 2d clipper poly2tri

Here is a quick description on how to generate 2D navmeshes from the level geometry to represent the paths an actor can take to get from one place to another.

The idea behind navmeshes is simple: Instead of building a graph of waypoints to represent the

First, we build a list of polygons of all the areas where an actor could potentially walk. In order for Clipper to properly merge them later, the polygons should overlap somewhat. In my case, floor tiles are always rectangular, so I simply enlarged their polygons by one (the smallest possible value since Clipper uses integer coordinates) in each direction.

We can then use Clipper to perform a union operation on the list to merge any overlapping polygons. Make sure to use

Now we need a list of polygons to represent the areas that are blocked by walls or any other entities. To generate each entity's polygon, take the vertices representing its 3D bounds (a bounding box, collision model or even the actual model), transform them to world space, project them onto the floor and construct their convex hull. The gift wrapping algorithm was fast enough for me and is easy to implement, but you may need a more sophisticated algorithm if your models have a significant number of vertices. Since I have very few entities that actually need a concave 2D outline, I decided that a convex hull is accurate enough in my case.

Since every point within the mesh should be a valid position for an actor, we need to expand the blocked areas to account for the size of the actors. Luckily, Clipper provides a very handy

Now that we have two sets of polygons, we can use Clipper to subtract the blocked areas from the walkable areas. The resulting list contains the modified walkable polygons as well as blocked polygons that are fully contained within a walkable polygon. These holes can be identified by their clockwise winding order. After this step, Clipper's work is done and we need to convert the polygons to Poly2Tri's representation.

Since we used the 3D bounds to generate the convex hulls, it is very likely that the polygons contain points that are in close proximity to each other. Since these points could lead to very long but thin triangles in the final mesh, it is wise to remove them before triangulating the polygons.

Having a mixture of very big and very small triangles in the final mesh makes it harder for a pathfinding algorithm to estimate the distances between nodes. By repeatedly subdividing the sides of a polygon until a desirable length has been reached, we can limit the maximum size of the triangles in the final mesh. Placing Steiner points allows even more control over the final shape of the triangles. These points can be placed at arbitrary positions within a polygon to subdivide it without altering its shape. In my case, with obstacles that are seldom smaller than one square meter/unit, I found that placing Steiner points in a grid every two meters/units produces nice results.

The polygons are now ready to be triangulated by Poly2Tri. While Clipper saves all polygons in a flat list and identifies holes by their opposite winding order, Poly2Tri works best when every walkable polygon is assigned a list of the holes it contains.

Triangulated mesh without and with additional Steiner points:

We now have a "triangle soup". However, in order to use it for pathfinding, each triangle needs to know about its neighbours.

We now have a list of triangles that we can use with A* or other pathfinding algorithms.

**paths**an actor can take to get from one place to another, we build a mesh that represents the**area**where an actor is allowed to walk. Since I'm quite happy with the way my implementation turned out, here is a quick description of how I generate my navmeshes from the level geometry in 2D using Clipper and Poly2Tri.## Define potentially walkable areas

First, we build a list of polygons of all the areas where an actor could potentially walk. In order for Clipper to properly merge them later, the polygons should overlap somewhat. In my case, floor tiles are always rectangular, so I simply enlarged their polygons by one (the smallest possible value since Clipper uses integer coordinates) in each direction.

## Merge walkable polygons

We can then use Clipper to perform a union operation on the list to merge any overlapping polygons. Make sure to use

`NonZero`as the`PolyFillType`for the operation. This tells Clipper to create polygons that are all wound the same way - counter-clockwise in my case. Using the default value (`EvenOdd`) may result in polygons that are wound the other way to represent holes.clipper.AddPolygons(walkablePolygons, PolyType.Clip) result := empty list clipper.Execute(ClipType.Union, result, PolyFillType.NonZero, PolyFillType.NonZero)

## Define blocked areas

Now we need a list of polygons to represent the areas that are blocked by walls or any other entities. To generate each entity's polygon, take the vertices representing its 3D bounds (a bounding box, collision model or even the actual model), transform them to world space, project them onto the floor and construct their convex hull. The gift wrapping algorithm was fast enough for me and is easy to implement, but you may need a more sophisticated algorithm if your models have a significant number of vertices. Since I have very few entities that actually need a concave 2D outline, I decided that a convex hull is accurate enough in my case.

blockedPolygons := empty list for each entity e: vertices := 3D bounds of e worldVertices := for each v in vertices => transform(v, e.WorldMatrix) floorVertices := for each v in worldVertices => [v.x, 0, v.z] hull := convexHull(floorVertices) add hull to blockedPolygons

## Expand polygons to account for actor size

Since every point within the mesh should be a valid position for an actor, we need to expand the blocked areas to account for the size of the actors. Luckily, Clipper provides a very handy

`OffsetPolygon`function to do just that. The method's`JoinType`parameter controls the number of vertices that Clipper inserts at acute angles. I found that using`Square`to chamfer sharp corners without inflating the vertex count too much produces nice results.for each polygon p in blockedPolygons: p := Clipper.OffsetPolygons(p, AMOUNT, JointType.Square)

## Subtract blocked areas from walkable areas

Now that we have two sets of polygons, we can use Clipper to subtract the blocked areas from the walkable areas. The resulting list contains the modified walkable polygons as well as blocked polygons that are fully contained within a walkable polygon. These holes can be identified by their clockwise winding order. After this step, Clipper's work is done and we need to convert the polygons to Poly2Tri's representation.

result := walkablePolygons for each polygon p in blockedPolygons: clipper := new clipper clipper.AddPolygons(result, PolyType.Subject) clipper.AddPolygon(p, PolyType.Clip) clear result clipper.Execute(ClipType.Difference, result) return result

## Optimize holes

Since we used the 3D bounds to generate the convex hulls, it is very likely that the polygons contain points that are in close proximity to each other. Since these points could lead to very long but thin triangles in the final mesh, it is wise to remove them before triangulating the polygons.

## Tessellate polygons

Having a mixture of very big and very small triangles in the final mesh makes it harder for a pathfinding algorithm to estimate the distances between nodes. By repeatedly subdividing the sides of a polygon until a desirable length has been reached, we can limit the maximum size of the triangles in the final mesh. Placing Steiner points allows even more control over the final shape of the triangles. These points can be placed at arbitrary positions within a polygon to subdivide it without altering its shape. In my case, with obstacles that are seldom smaller than one square meter/unit, I found that placing Steiner points in a grid every two meters/units produces nice results.

## Triangulate

The polygons are now ready to be triangulated by Poly2Tri. While Clipper saves all polygons in a flat list and identifies holes by their opposite winding order, Poly2Tri works best when every walkable polygon is assigned a list of the holes it contains.

walkablePolygons := all polygons with counter-clockwise winding order blockedPolygons := all polygons with clockwise winding order triangleList := empty for each w in walkablePolygons: w.Holes := every polygon b in blockedPolygons that is contained in w triangulate(w) add every triangle in w.Triangles to triangleList return triangleList

Triangulated mesh without and with additional Steiner points:

## Find neighbours

We now have a "triangle soup". However, in order to use it for pathfinding, each triangle needs to know about its neighbours.

triangleIDs := empty map from polygon to int vertexIDs := empty map from vertex position to int vertices := empty list // give every triangle and vertex a unique ID for each triangle t in triangleSoup: triangleIDs[t] := triangleIDs.Count for each vertex p in t: if p not in vertexIDs: vertexIDs[p] = vertexIDs.Count add p to vertices edge { V1, V2 } // undirected edge, edge1 is equal to edge2 if they contain the same vertices // group all triangles that share an edge edgeToTriangles := empty map from edge to list of triangles for each triangle t in triangleSoup: edges := list of the three edges of t for each edge e in edges: add t to edgeToTriangles[e] // gather each triangle's neighbours for each triangle t in triangleSoup: myID := triangleIDs[t] edges := list of the three edges of t for each edge e in edges: for each triangle n in edgeToTriangles[e]: add n to t.Neighbours

We now have a list of triangles that we can use with A* or other pathfinding algorithms.

## About the Author(s)

*Christoph Romstöck is a student of computer science at the University of Erlangen-Nürnberg (FAU) with a passion for compilers, 3D graphics and minimally short "about the author" sections.*

## License

*GDOL (Gamedev.net Open License)*

## 4 Comments

I used the clipper library in the context of 3d navmeshes as well. Not in terms of generating the mesh, but in handling dynamic obstacles from the base navmesh regions.

Rather than using triangles though I use http://code.google.com/p/polypartition/ to generate optimal convex regions from the results.

It would be nice to have everything in a single library, in this case i'd be willing to participate (sp??) with my delaunay triangulator and voronoi generator.

What do you think guys ?

Interesting ... i'm using these tools to generate geometry within a level builder.

If anyone out there is generating nav meshes for a 3D world, without a doubt the best freely available solution is Recast and Detour written by ex Crytek engineer Mikko Mononen. Which comes with both runtime navigation mesh logic and pathing as well as a robust navmesh generator. It works by creating a voxel mold of your polygon soup and then using various filters to triangulate the mesh. It even comes with a great ui tool to preview the results and path finding. Well worth looking into.

https://github.com/memononen/recastnavigation

http://digestingduck.blogspot.ca

(creators blog)

i think it's good idea if the mesh small, but the number of triangle grow bigger, the algorithm O(n^2) would be very slow. i tried and got 10fps with path finding on 100x(50 vertex polygon).

here another aproach which reduce the number of node(vertex), http://www.david-gouveia.com/portfolio/pathfinding-on-a-2d-polygonal-map/

Note: GameDev.net moderates article comments.