Comparisons between two samples with multiple endpoints are often encountered in many real applications and Hotelling's T^2 test (HT) may suffer from loss of efficiency when multivariate normality assumption is violated. To overcome this issue, we propose a group Hotelling's T^2 test (GHT) where HT is conducted within each group after inverse normal transformation and then use the maximum value among combined statistics based on $p$-values at the group-level. Extensive simulations show that GHT is more robust than HT and some other existing procedures. Finally, the applications to plasma-renin activity in serum study and the ageing human brain further demonstrate the performance of GHT.