I suspect you're going to find that you can't get enough resolution out of a 10g accelerometer to accomplish what you want to do.
The obvious answer is to mount the sensor in the robot, and start making runs, logging the sensor data to something. Induce turns in the robot, and see what output you get. If you notice a correlation between the robot turning and sideways acceleration, then you can probably write something simple that provides feedback on how much the sensor believes the vehicle is going off course.
Of course, its much simpler to use odometry encoders to go in a straight line...