Baxter Gazebo simulator results ten-times worse than real robot

[masamouta1398]

Dear All

I am running the simulation on the Gazebo7. I can see the controller is learning but the error(MAE) is around 0.1 after several trials which is almost ten times worse than the results from the real robot that the authors have reported in their paper. Does anyone have any idea why the simulation result is so bad and the robot end-effector does not go through a nice circular or eight-like trajectory after learning process is done?