collision_avoidance_env.py line 356 about rl_collision_avoidance HOT 3 CLOSED

Another question is

rl_collision_avoidance/ gym-collision-avoidance/ gym_collision_avoidance/ envs/ policies/ CADRL/ scripts /multi /gen_rand_testcases.py line 214

Could you please explain why we need to make sure "straight line solution is not permitted"?

from rl_collision_avoidance.

Hi @TUzizi - good catch about the reward function... intuitively, + makes more sense because you'd want the penalty to decay to zero as the distance between agents gets sufficiently large. But, I think there might be a typo in some of the papers, since the original ICRA17 paper uses - just as this code implements, but the IROS18 and Access21 use a +. I'm not 100% sure about this anymore.

For the random test case code, it's very old code that I don't suggest reading too much into. It was written ad-hoc to create some scenarios for training/testing with sufficient interaction between agents to contain useful learning signals, but sometimes also let the agents have a simple straight line path to the goal (since that's a valid real-world scenario sometimes). I think the latter is the legacy reason for the line you ask about. There are probably a lot of better ways of creating such scenarios, and maybe even cool research ideas with curriculum learning as a more principled approach.

from rl_collision_avoidance.

Got it. Thank you.

from rl_collision_avoidance.