Enhancements

The graph generation procedure we've discussed works well enough for explaining the algorithm, but there are many ways in which it can be improved. These enhancements make the algorithm use less CPU and memory, and allow it to handle more cases. Let's take a look at a few.

Obstacles that touch

Perhaps you picked up on it—none of the circular obstacles in the examples given so far have overlapped or even touched. Allowing circles to touch makes the pathfinding problem a little, but not much, harder.

Bitangents

Recall that bitangents can be found using this formula for internal bitangents: \[\theta = \arccos{{r_A+r_B}\over{d}}\] and this one for external bitangents: \[\theta = \arccos{{\lvert r_A - r_B \rvert} \over d}\]

When two circles touch or overlap, there are no internal bitangents between them. In this case \({r_A+r_B}\over d\) is greater than one. Since arccosine is undefined for inputs outside its domain of \([-1, 1]\), it's important to check for circle overlap before performing the arccosine.

Likewise, if one circle completely encloses the other, then there are no external bitangents between the circles. In this case \({r_A - r_B} \over d\) is outside the range \([-1, 1]\), and it has no arccosine.

Surfing edge line of sight

When obstacles are allowed to touch or overlap, new cases arise in calculating surfing edge line of sight. Recall the calculation of \(u\), the fraction of distance along the surfing edge at which a perpendicular ascender to the point touches the edge. When circles are not allowed to touch, a value of \(u\) outside \([0,1]\) means that the circle cannot touch the edge because to do so it would have to touch one of the endpoints of the edge. This is impossible, because the endpoints of the edge are already tangent to (and thus touching) other circles.

However, if circles are allowed to overlap, then values of \(u\) outside \([0,1]\) might block line of sight along the edge. This corresponds to cases where the circle is off the end of the surfing edge, but covering or touching an endpoint. To catch these cases mathematically, we clamp \(u\) to the range \([0,1]\): \[E = A + clamp(u, 0, 1) * (B - A)\]

Hugging edge line of sight

When obstacles are allowed to touch or overlap, hugging edges can be blocked by obstacles just as surfing edges can. Consider the hugging edge in the diagram below. If another obstacle touches the hugging edge, it’s blocked and should be thrown out.

To determine whether a hugging edge is blocked by another obstacle, use the following method to determine the points at which the two circles intersect. Given circles centered on points \(A\) and \(B\) with radii \(r_A\) and \(r_B\), where \(d\) is the distance between \(A\) and \(B\), there are a few cases to check for first. If the circles are not touching (that is, \(d > r_A + r_B\)), if one circle is inside the other (\(d

If none of these cases hold, then the circles intersect at two points—if the circles are tangent to each other, these two points are coincident. Consider the radical line which connects the intersection points; it's perpendicular to the line connecting \(A\) and \(B\) at some point \(C\). We can calculate the distance \(a\) from \(A\) to \(C\) as follows: \[a = {r_A^2 - r_B^2 + d^2 \over 2d } \] Having found \(a\), we can find the angle \(\theta\): \[\theta = \arccos {a \over r_A} \] If \(\theta\) is zero, the circles are tangent at \(C\). Otherwise, there are two intersection points, corresponding to positive and negative \(\theta\).

Next, determine whether either of the intersection points fall between the start and end points of the hugging edge. If this is the case, then the obstacle blocks the hugging edge, and we should discard the edge. Note that we don’t have to worry about the case where the hugging edge is entirely contained within an obstacle, as the line of sight culling for surfing edges will have already thrown the edge out.

After making the changes for bitangent calculation and line of sight for surfing and hugging edges, everything else just works.

Variable actor radius via Minkowski expansion

When navigating a round object through a world of round obstacles, we can make a few observations that simplify the problem. First, we can make things easier by noticing that moving a circle of radius r through a forest of round obstacles is identical to moving a point through that same forest with one change: each obstacle has its radius increased by r. This is an extremely simple application of Minkowski addition. If it has a radius larger than zero, we’ll just increase the size of the obstacles before we start.

Delayed edge generation

In general, the graph for a forest of \(n\) obstacles contains \(O(n^2)\) surfing edges, but since each of these must be checked for line of sight against \(n\) obstacles, the total time to generate the graph is \(O(n^3)\). Additionally, pairs of surfing edges can induce hugging edges, and each of these must be checked against every obstacle for line of sight. However, because the A* algorithm is so efficient, it normally looks at only a fraction of this large graph to produce an optimal path.

We can save time by generating small portions of the graph on the fly as we proceed through the A* algorithm, rather than doing all of the work up front. If A* finds a path quickly, we'll generate only a small part of the graph. We do this by moving edge generation to the neighbors() function.

There are several cases. At the beginning of the algorithm, we need the neighbors of the start point. These are surfing edges from the start point to the left and right edges of each obstacle.

The next case is when A* has just arrived at point \(p\) on on the edge of obstacle \(C\) along a surfing edge: neighbors() needs to return the hugging edges leading from \(p\). To do this determine which surfing edges leave the obstacle by computing the bitangents between \(C\) and every other obstacle, throwing away any that do not have line of sight. Then find all the hugging edges that connect \(p\) to these surfing edges, discarding those that are blocked by other obstacles. Return all of these hugging edges, saving the surfing edges for return in a subsequent neighbors() call.

The last case is when A* has traversed a hugging edge along obstacle \(C\) and needs to leave the \(C\) via a surfing edge. Because the previous step calculated and saved all the surfing edges, the correct set of edges can just be looked up and returned.

Cull cusped hugging edges

Hugging edges connect surfing edges which touch the same circle, but it turns out that many of these hugging edges are not eligible to be used in any optimal path. We can speed up the algorithm by eliminating them.

An optimal path through the forest of obstacles always consists of alternating surfing and hugging edges. Suppose we're entering at node \(A\) and are trying to decide how to exit:

Entering through \(A\) means we're going clockwise\(\circlearrowright\). We must exit through a node that keeps us going clockwise\(\circlearrowright\), so we can only exit through node \(B\) or \(D\). Exiting through \(C\) creates a cusp\(\curlywedge\) in the path, which will never be optimal. We want to filter out these cusped edges.

First note that A* already treats each undirected edge \(P \longleftrightarrow Q\) as two directed edges, \(P \longrightarrow Q\) and \(Q \longrightarrow P\). We can take advantage of this by annotating the edges and nodes with directions.

1. The nodes \(P\) become nodes with a direction, either clockwise \(P\circlearrowright\) or counterclockwise \(P\circlearrowleft\).
2. The undirected surfing edges \(P \longleftrightarrow Q\) become directed edges \(P,p \longrightarrow Q,\hat{q}\) and \(Q,q \longrightarrow P,\hat{p}\), where \(p\) and \(q\) are directions, and \(\hat{x}\) means the opposite direction of \(x\).
3. The undirected hugging edges \(P \longleftrightarrow Q\) become directed edges \(P\circlearrowright \longrightarrow Q\circlearrowright\) and \(P\circlearrowleft \longrightarrow Q\circlearrowleft\). This is where the filtering happens: we don't include \(P\circlearrowright \longrightarrow Q\circlearrowleft\) and \(P\circlearrowleft \longrightarrow Q\circlearrowright\), because changing direction introduces cusps\(\curlywedge\).

In our diagram, node \(A\) would become two nodes, \(A\circlearrowright\) and \(A\circlearrowleft\), and have an incoming surfing edge \(\longrightarrow A\circlearrowright\) and an outgoing surfing edge \(A\circlearrowleft \longrightarrow\). If the path entered through \(A\circlearrowright\) then it must exit through a \(\circlearrowright\) node, which would be either the \(B\circlearrowright \longrightarrow\) surfing edge (via the \(A\circlearrowright \longrightarrow B\circlearrowright\) hugging edge) or the \(D\circlearrowright \longrightarrow\) surfing edge (via the \(A\circlearrowright \longrightarrow D\circlearrowright\) hugging edge). It can't leave through \(C\circlearrowleft \longrightarrow\) because it changes rotation direction, and we have filtered out the \(A\circlearrowright \longrightarrow C\circlearrowleft\) hugging edge.

By filtering these cusped hugging edges out of the graph, we make the algorithm more efficient.

Crossing edge culling

Cull partial paths whose final surfing edge crosses the penultimate surfing edge.

Polygonal obstacles

See Game Programming Gems 2, Chapter 3.10, Optimizing Points-of-Visibility Pathfinding by Thomas Young. It covers the node culling but for polygons instead of circles.

References