Learning from Sparse Demonstrations