Hi Anna,
good catch, that is definitely important to note as it means that the benchmark results may not be exactly replicable after the refactor (the 4th bullet point Tristan mentioned here). From what I can see from a brief look at the use of self.rng in the EnvironmentDataLoader it looks like it is only used to be passed forward to specific transforms (which have needs_rng==True). The transforms that need a random number generator are those that add random noise to the observations for testing, so if you use a benchmark experiment that does not add noise (e.g., base_config_10distinctobj_dist_agent or base_77obj_surf_agent), the before and after results should still be identical.
Actually, even the experiments that add noise after sensor processing in the SM should still be the same (like randrot_noise_77obj_surf_agent) as those don’t add the noise in the transform. So thinking about it a bit more, only the randomrot_rawnoise_10distinctobj_surf_agent experiment might be affected by this.
Best wishes,
Viviane