MultiAgent RGB Observation

[gauravkuppa]

When I try to run python3 multiagent.py --num_drones 4 --obs rgb, I get the error [ERROR] ObservationType.RGB for multi-agent systems not yet implemented. However, in the results from the README, I can see that you tested multiple drones with vision. Is this feature available?

	Lenovo P52 (i7-8850H/Quadro P2000)	2020 MacBook Pro (i7-1068NG7)
Rendering	OpenGL	CPU-based TinyRenderer
Single drone, no vision	15.5x	16.8x
Single drone with vision	10.8x	1.3x
Multi-drone (10), no vision	2.1x	2.3x
Multi-drone (5) with vision	2.5x	0.2x
80 drones in 4 env, no vision	0.8x	0.95x

If not, I am willing to contribute to this repo to add this functionality. What is the status of this work?

[JacopoPan]

Hi @gauravkuppa

you can already use ObservationType.RGB with BaseMultiagentAviary and all subclasses. What is not implemented yet is a learning script (like multiagent.py) with an RL agent intended to handle that type of observation. Does this answer your question?

[gauravkuppa]

Could you point me towards how to implement multiagent learning?

[JacopoPan]

Do you have a MARL approach in mind? The multiagent.py is meant as an example leveraging the central critic/decentralized polices example in RLlib but that's a very specific and arbitrary choice.