Is there a standard way to combine the inverted control of ExternalEnv with the keyed policy dict of MultiAgentEnv, or does this need to be a custom Environment?
Hey @Eric_Adlam , yes, this is possible today. There is a ExternalMultiAgentEnv class that inherits from ExternalEnv that handles multi-agents dicts (observations, actions, rewards, dones dicts).
You can take a look at our Unity3D client/server example here, where we use this kind of setup:
ray/rllib/examples/serving/unity3d_[client|server].py
1 Like