Map reconstruction for autonomous mobile robots navigation needs to deal with uncertainties, imprecisions and even imperfections due to the limited sensors quality and knowledge acquisition. An improved fusion machine is proposed by replacing the classical conjunctive operator with T-norm operator in Dezert-Smarandache Theory (DSmT) framework for building grid map using noisy sonar measurements. An experiment using a Pioneer II mobile robot with 16 sonar detectors onboard is done in a small indoor environment, and a 2D Map is built online with our selfdeveloping software platform. A quantitative comparison of the results of our new method for map reconstruction with respect to the classical fusion machine is proposed. We show how the new approach developed in this work outperforms the classical one.