Vector-env workers: reset(seed=) does not cover constructor-time np.random draws
In RL vector envs (AsyncVectorEnv / SubprocVecEnv style), the env_fn runs inside the worker, so any global np.random draw in an env's __init__ (random map generation, wind/turbulence phase, domain-randomization parameters) happens after fork, or after the 3.13.8+/3.14.1+ forkserver preload, and is identical in every worker. Passing distinct seeds to reset(seed=...) does not fix it: that reseeds only the env's own self.np_random, and the constructor already ran. When auditing, read the constructors and wrappers that run in the worker, not just step/reset, and check a parameter the env never redraws (get_attr it from each worker). A related pattern: code that reseeds torch per worker ('otherwise each worker generates same experience') but leaves an OU / Gaussian exploration noise class on np.random. If the parent draws nothing between rounds, every round's pool starts from the same state, so the 'noise' is one fixed series for the whole run. To isolate additive exploration noise, run the real collection code with a fixed policy (mean 0, std 1e-12): the actions are then the noise.
Reproduced with gymnasium 1.4.0's AsyncVectorEnv and NumPy 2.5.4, using a toy env whose __init__ sets self.wind = np.random.random(). The parent script loads numpy.random at top level, and I read the value back with get_attr('wind') after reset(seed=[1, 2, 3, 4]). CPython 3.12.16: context='fork' gives 1 distinct of 4; 'spawn' and 'forkserver' give 4/4. CPython 3.14.10: 'forkserver' run as a script file gives 1/4, but the same file run with 'python -m' gives 4/4, because the preload happens only for a script-file __main__. 'spawn' gives 4/4. self.np_random was distinct in every case. So spawn really does avoid this constructor-time duplication, unlike the pickled-RandomState case. I did not test the OU-noise paragraph.