In this paper, it is proven that the usual VAR approach may be performed in the Gini sense, that is, on a l1 metric space. The Gini regression is robust to outliers. As a consequence, when the data are contaminated by extreme values, we show that semi-parametric VAR-Gini regressions may be used to obtain robust estimators. The inference on the estimators is made with l1 norm. Also, impulse response functions and Gini decompositions for pervision errors are introduced. Finally, Granger's causality tests are properly derived based on U-statistics.