Carlos Luis

> it seems that the failures come from `pygame` not being installed somehow, probably an issue from gym... Only a subset of the failures, those are fixed now.

@RedTachyon I'm currently investigating a bug which may be related to https://github.com/openai/gym/pull/2422. Here's a stack trace of the type of errors: ``` tests/test_save_load.py:216: _ _ _ _ _ _ _...

> I think I found the issue - the custom RNG class inherits from the numpy Generator for compatibility purposes, but when it gets pickled/unpickled, it defaults to the numpy...

Update on this: - Confirmed that https://github.com/openai/gym/pull/2639 solves the bug I mentioned before, so in the next Gym release those tests will pass. - There's three remaining tests failing, two...

Regarding the "mean reward below threshold" problem, after some further investigations the root cause is the change in seeding behaviour in gym. SB3 sets seeds here: https://github.com/DLR-RM/stable-baselines3/blob/e88eb1c9ca98650f802409e5845e952c39be9e76/stable_baselines3/common/base_class.py#L576 Before gym 0.22.0,...

Similarly, the HER test failure was also caused by the change in RNG. The failing test required that after training for 200 steps, the env had to be left in...

@araffin I would have to check the warnings, but now every time env.seed is called will raise a warning. I reverted the threshold back to 90 and changing the seed...

Further investigations on these: @RedTachyon is right, the determinism in the test (both using gym v0.21.0 and current master) actually comes from setting the seed on the action space rather...

> @carlosluis i think there was a misunderstanding with my remark here: [#780 (comment)](https://github.com/DLR-RM/stable-baselines3/pull/780#issuecomment-1062235112) > > please do not change `VecEnv` signature but make `VecEnv` `seed()` method forward compatible (so...

> > AFAIK, to make VecEnv seed() forward compatible we should internally call reset(seed=seed) for every env, similar to how it's now done in DummyVecEnv. > > yes > >...