The following applies to DDPG and TD3, and possibly other models. The following libraries were installed in a virtual environment:

numpy==1.16.4
stable-baselines==2.10.0
gym==0.14.0
tensorflow==1.14.0

Episode rewards do not seem to be updated in model.learn() before callback.on_step(). Depending on which callback.locals variable is used, this means that:

• episode rewards may n

I'll post it as a question as I am not quite sure that it is a bug. I have been experimenting for a while with the library in a custom environment for a school project and I am really interested in the reproducibility of the result. I have read the disclaimer in the documentation that reads that reproducible results are not guaranteed across multiple platforms or different versions of Pytorch. Ho

There seem to be some vulnerabilities in our code that might fail easily. I suggest adding more unit tests for the following:

• Custom agents (there's only VPG and PPO on CartPole-v0 as of now. We should preferably add more to cover discrete-offpolicy, continuous-offpolicy and continuous-onpolicy)
• Evaluation for the Bandits and Classical agents
• Testing of convergence of agents as proposed i

This is more or less similar to what we were doing long time ago with pybullet (#110, #149).

In few words, after #346 the pure-Python gym-ignition package will depend on scenario. When gym-ignition gets installed in a system (currently only Ubuntu is supported), the wheel of scenario will be installed first. Since the wheel of scenario is not self-contained (i.e. it needs to find in t