Nice visualizations and explanation! You might want to make it clearer that the agents don't actually receive any visual observations, but rather directly the

Nice visualizations and explanation! You might want to make it clearer that the agents don't actually receive any visual observations, but rather directly the xy positions of all other agents and objects. This also seems very similar to "Capture the Flag: the emergence of complex cooperative agents" (https://deepmind.com/blog/article/capture-the-flag-science https://deepmind.com/blog/article/capture-the-flag-science)? Regarding the conclusion: > We’ve provided evidence that human-relevant strategies and skills, far more complex than the seed game dynamics and environment, can emerge from multi-agent competition and standard reinforcement learning algorithms at scale. These results inspire confidence that in a more open-ended and diverse environment, multi-agent dynamics could lead to extremely complex and human-relevant behavior. This has been well established for a while already, e.g. the DeepMind Capture the Flag paper above, AlphaGo discovering the history of Go openings and techniques as it learns from playing itself, AlphaZero doing the same for chess, etc.