Your idea reminds me a bit of artificial potential fields (they've been used for robot control for a long time). It sounds like your idea could work really well for your problem. One thing to keep in mind though: You are assuming that the control output (speed + direction) only depends on the robot's current position (which grid cell it's in). In reality, the control output probably depends on more inputs: the robot's position, orientation, steering angle and velocity.

I'm not aware of existing implementations of that idea, but there may very well be some.

Oh, one more thing: your grid map isn't really a planning method; it's Memoization. You will still need a planning (or learning) method to fill in the grid map. Perhaps SBPL can be used for that; the library's authors have used it for the DARPA urban challenge before (see SBPL publications). But perhaps you'll want to try something simpler first.