Question Why decision_steps in unityenvironment suddenly get nothing?

I try to use custom PPO algorithm to train a continuous action env(3DBall or others) in python by ml-agents env, just like tutorial https://github.com/Unity-Technologi...op/colab/Colab_UnityEnvironment_2_Train.ipynb, which use custom Qlearning to train discrete action space env.

At the beginning 3DBall env can run normally, but several seconds later it terminates.


I find decision_steps suddenly get nothing from last env step(it works normally at the beginning and the obs shape is [num_agents, obs_dim]), and I pass the obs to net to get the action(actually null), so compiler report an error above.




Any ideas will be appreciated. Thanks!