has anyone successfully trained multi agent problem : flockaviary? about gym-pybullet-drones

[JacopoPan]

is this a problem with environment (flockaviary.py) or the algorithm (multiagent.py)

(there's something to say about action/observation types and the policy architectures but) the short answer is both: the reward function in flockaviary.py is the one you mention, not one necessarily intended for flocking, and multiagent.py is still the least mature piece of code in this repo (e.g., it needs work to extend to more than 2 drones)

you can use test_multiagent.py to have a look at what the trained policies are (and maybe infer where it is going wrong)

[JacopoPan]

I realized that test_multiagent.py was not working properly because the input to compute_single_action() was not in the correct order. The default execution of script multiagent.py (a simple 1D leader-follower example) can now be trained and tested in ~10'.