We present a deep learning framework to reconstruct the three-dimensional cosmic dark matter density field from observed halo distributions. By training a feed-forward neural network (FNN) on the Millennium Simulation, we establish a nonlinear mapping from the halo group catalogs to the underlying dark matter density. At spatial resolutions of (10, 5, 2) h−1 Mpc, our method significantly outperforms the standard halo cloud-in-cell approach, especially in low-density regions where halo tracers are sparse. Our scheme could recover both the density probability distribution function and the power spectrum statistics as well. The FNN achieves reconstruction errors of less than half an order of magnitude across the entire density dynamic range, demonstrating robust generalization across cosmic environments. These results establish a foundation for sparse galaxy tracers of the real dark matter density field in the forthcoming large-scale galaxy surveys (DESI, LSST, and Euclid).