Most conventional CAM algorithms are geometry-based. They create toolpaths that are usually parallel to either the coordinate axes ('zigzag'-paths) or to the part contour (contour-parallel, or spiral paths). One problem with this geometry-based stuff is that you don't take into account the cutting forces. The algorithm has no idea about how much material is removed while the tool is moving. A quick patch is to run a cutting simulation after the path is created and adapt the feedrate for constant material removal rate (MRR).
Another option is to base the toolpath algorithm on a stock model. Then you know the shape of the stock at all times and you can control the MRR or cutter engagement angle. To quickly test how this could work I made a small test in matlab.
The cutter (green circle) is moved around by some rules, and cut's the red pixels as it travels over them. Cut pixels are drawn in blue. The cutter is moved around in discrete steps in some direction, and you're only allowed to cut a certain number of pixels per move. The tricky part is coming up with the rules for our 'lawn-mower' robot. Now I'm using a simple idea: If the past move was made at an angle alfa, try to take the next step in the same direction, but if that's not possible increase alfa until the MRR goes down to some preset value.
This idea will need refinement so that the robot can cope with walls, can do cutting in only one direction (climb vs. conventional) etc. etc., but this seems like a promising and fun start!