Motivation
We currently use a custom function get_stats_random_rollout for stats computation in the examples.
We should use ObservationNorm.init_stats instead.
We have 2 options:
- Rewrite
get_stats_random_rollout from within to just call the transform method
- Replace every call to
get_stats_random_rollout() by a call to ObservationNorm.init_stats.
The first solution may be more readable but adds one level of coding that the second does not have, and perhaps has little advantages except for conciseness of the training script.