andnp / pyrlenvs Goto Github PK

AI gym has changed lots of their code to using float32 precision for their outward facing API to comply with their box constraints. However, internally they are still often using float64 in their physics integrators. Worse, they are often using float32 for the first integration step and implicitly converting this to float64 for the remaining integration steps. This is bad.

This accumulation of floating point error means I cannot do Monte Carlo rollout compliance testing any longer, because I do not want to match gym's weird choice. Changing this means accepting that I will deviate from Gym environments though, which should absolutely come with a warning.