Minimal Interface for Gymnasium, an RL environment repository which replaces OpenAI Gym. Might implement compatibility layer for CommonRLInterface or MinimalRLCore if there is interest.
make(id::String;
unwrap=false, # whether the environment is returned with the default wrappers or unwrapped
render_mode=nothing, # corresponds to the render_mode in the gymnasium environment
max_episode_steps=nothing,
autoreset=false,
disable_env_checker=nothing,
kwargs... # Untested, but technically should send optional kwargs to environment constructor in python.
)
make(name::String, version::Int; kwargs...) # calls the above functionThen the interface is juliafied in the usual ways:
step!(env::GymnasiumEnv, action)
reset!(env::GymnasiumEnv{T}; seed::Union{Nothing, Int}=nothing, options::Union{Nothing, Dict}=nothing) where {T}
render(env::GymnasiumEnv)
close!(env::GymnasiumEnv)I have not tested on every environment. To add compatibility for an observation type that I don't implement convert_obs similar to the ndarray function:
convert_obs(::Val{:ndarray}, pyobs::Py) = pyconvert(Vector, pyobs)where :ndarray comes from the call of convert_obs in the GymnasiumEnv wrapper class:
function GymnasiumEnv(id::String, pyenv::Py)
obs, info = pyenv.reset() # reset to get the obs type
env = GymnasiumEnv(pyenv, id, convert_obs(obs), false, 0.0, info)
return env
end
function convert_obs(pyobs::Py)
t_str = pyconvert(String, @py type(pyobs).__name__)
convert_obs(Val(Symbol(t_str)), pyobs)
endThe info dict is returned as a PythonCall.Py object, and the reward and terminal values are assumed to be Float64 and Bool respectively.