I guess you will need to. Actually I think ( I am not 100% sure ) you should cast a circle (or a box) which is slightly more complicated. What you do in this case however, is enlarge the walls by the agent radius obtaining a sort of rounded walls (the corners of the inflated walls will be filled with arcs), and then cast a ray against them; the ray shall have length = distance_of_wall_detection (which could be current_velocity * time_of_prediction). You will need segment to segment intersections and segment to circle intersection ( http://www.gamedev.n...e-intersection/ ) By the way, (I am saying this just not to let anything out) do you connect the node of the graph based only on the criterion that two adjacent cells are passable? If yes you should instead cast a box as large as a gridcell (or as large as an agent bounding box) towards an adjacent cell and connect them only if no intersection with any static obstacle occurred. In this way the smoothed path you showed might not be computed: the agent will go up to the cell with the value = 2 and the steer by 90° towards the goal. I hope this helps.
Are these walls defined when an agent computes the path?
Are they defining the maximum distance from the path that an agent must follow (path with a width)?
you shift the path segments alongside their normal (both ways) and then change the extreme of these segments using the intersection with the lines passing through the extremes of the adjacent segments in order to roughly fill the gaps that creates (like the picture I posted before).
Does this help?
Maybe you want the agents to follow a path without getting too far from the segments of the path that connects the nodes, is that it? The example you posted shows a behaviour that seems just to perform a clamping of the agents positions against walls. If you need the agents to follow the path having their distance from the current path segment be less than an amount, you need just to define which is the current path segment and the distance from the agent. Compute the closest node to the agent find the graph links _in the path_ that share that node have the distance of the agent be less than an amount for one or both of the graph link just found and beware of "corners" as shown in this figure http://i.imgur.com/jmpPK.jpg (the red lines are the path segments normals, the blue lines are the imaginary walls that define your path width, the red ugly circles are the wrong agents positions, as you can see those agents in the corners shall compute their distance from the walls carefully) Just kindly check this strategy because I am not 100% sure it will find the correct path segments from which to measure the agent distance.
I guess yes they are freeform approximated by the triangles of a mesh, those roads were modeled on top of a google map fragment. Should I choose a semi-automatic approach? That is once the graph is generated I build a simple GUI to edit the nodes, so that I can add and smooth curves or simplify unnecessary complexity and so forth...
I need also to understand if the graph should represent the actual lanes of the roads (that could be smoothed on run time by a path follower system) or if it should just be a sort of high level graph and then the agents will figure out what the lane to follow is.