We present a novel, data-driven kinodynamic motion planner. Our sampling-based planner is based on using a physics simulator as a black box to compute a trajectory considering dynamics, even when we cannot derive exact propagation functions. To improve its overall efficiency, we pre-compute a motion database containing different motions simulated with different controls and states defined in the local frame of a robot. We then use the motion database to efficiently estimate the simulated trajectory during iterations of our planner. When the planner requests the best control to reach a desired state from a query state, we retrieve nearby motions that are close to the query state and pick the motion that is closest to the desired state for the tree extension. To control accuracy of our planner with a high efficiency, we lazily validate retrieved motions. The pre-constructed motion database contains modular trajectories and thus can be reused for other test cases, where we have different composition of obstacles or different start/goal states.