a certain chance to fail

[yueyang130]

Hi, @denisyarats , Thanks for your great work!

I rerun your code and find there is a certain chance to fail to train a good agent. For example, as the below picture shows, I run Cartpole Swing 4 seeds with your config file except that I shorten the total frame to 200K. According to curves in the paper, 200K frames are enough for agents to perform well in Cartpole Swing. However, one of 4 runnings failed. I have checked data in curves folder . It seems that you don't encounter this problem.

Similar problems occur in other tasks such as Walker Run. I just rerun these two tasks so far. But I doubt this problem would happen also in other tasks.
Now I doubt the version of pytoch or Mujoco caused this since I used my own docker. I will try to use the same version with you and report the new results later. If you have encountered similar problems or have any idea about why this happened, please tell me.