The same level of agents in the paper cannot be reproduced from the source code

[leoranlmia]

Dear authors:
Thanks for sharing this code! This is a great work! However, when I try to train the robust agents with PGD solver, I cannot get the same level of agents of the paper.
After testing these agents, only get:

• 0.0 +- 0.0 average reward in the Bankheist environment.
  
• 0.0 +- 0.0 average reward in the RoadRunner environment.
  
• 21.46 +- 1.6150541786577934 average reward in the Freeway environment.
  
• -21.0 +- 0.0 average reward in the Pong environment.
  

Maybe my configured environment is the cause? Could you provide the exact version of python and main packages (torch, gym, numpy etc.)?
Or what else might be the cause？

Looking forward to hearing from you! Thank you!