CAD to Machine Learning

Hello Ladys and Gentlemen,
for my Thesis i´m looking for a proper way to train an AI (Reinforcement Learning Algorithm) on a in CAD modelled Robot. I did some research about how could that task be solved and found a few solutions. My way at the moment is to build a model in Fusion360 and export an .sdf file (and .stl files) with the add-on SDFusion. Then i simulate the modell in PyBullet to check if the behavior is as expected. I would say it isnt… the Inertias are pretty crappy (for example i_zy = 6.542*10^-45 on a symmetric modell), the axis of rotation are also not always the way i defined them and not even all joints can be moved (limits are set <0 or >0!). If anyone has done some kind of work i would appreciate if we could get into a discussion. Or if somebody has an idea of what i could maybe doing wrong, please tell me.

Thanks for your help