UnityGymWrapper.jl

Unity ML-Agents (Gym) wrapper using PyCall.

JuliaRL Implementations

Requirements (Python Environment)

• Unity ML-Agents - Release 2
• Unity ML-Agents Python Interface (Envs) - v0.16.1
• Unity ML-Agents Gym Wrapper - v0.16.1 (same limitations)
• Environment Executables

API

• Create/ close environment:

Environment(path; nographics=true, usevisual=false, uint8visual=false, multipleobs=false, kwargs...)

# logFile path to prevent unity terminal output
env = Environment("envs/Basic", logFile=pwd()*"/envs/logs/Basic.log")
close!(env)

• Environment interaction

s = reset!(env) # needs to be called after (done == true) step
s, r, done, info = step!(env) # random action
s, r, done, info = step!(env, action) # eltype(action) == eltype(env.actionspace)

• Environment data (check examples & tests)

# Required action array size & eltype (model output)
size(env.actionspace)
length(env.actionspace)
eltype(env.actionspace)

sample(env.actionspace) # random action

# Discrete/ MultiDiscrete actionspace
# Max ranges for each dim
actiondim1 = env.actionspace.actions[1]
actiondim2 = env.actionspace.actions[2]
...
# Box actionspace
# Range tuples for each dim
l, h = env.actionspace.actions[1]
...

# State/ Observation size & eltype (model input, single or multiple observations)
env.observationspace::Union{ObservationSpace, Vector{ObservationSpace}}
size(observationspace)
length(observationspace)
eltype(observationspace)
l, h = env.actionspace.vals[1]