Agents' abnormal behaviors about gym-collision-avoidance HOT 5 CLOSED

Thanks for sharing this behavior and screenshots! I am not too surprised by this behavior, since the subgoals are much further away than what was normally seen during training. Also, the CADRL policy was trained in "smaller" scenarios compared to GA3C-CADRL, so that makes sense to me as well. In practice, we treat these policies as local planners, so we usually have a global planner that takes goals at the range of ~25m and regularly (~10Hz) provides a local subgoal ~5-10m in front of the vehicle that should lead toward the longer goal. Alternatively, one could re-train the policies in larger environments, but that's not something we tried much.

from gym-collision-avoidance.

Thank you so much for the reply and explantions; that makes a lot of sense to me! Would you possibly able to point me to the part of the code in the repo where I can gain more understanding about the local planner and/or global planner?

from gym-collision-avoidance.

There isn’t really any code related to that distinction here in this repo. The cadrl_ros repo is an example of a local planner that wraps around the ga3c-cadrl policy. Bruno Briton’s 2021 RAL paper talks about one way for handling the interface btwn global and local planners, for example

from gym-collision-avoidance.

https://arxiv.org/abs/2102.13073

from gym-collision-avoidance.

Gotcha! Thank you so much for your help! I actually do have one more unrelated question. But to keep a cleaner format, I will open another request. I hope my continuous questions don't bother you too much!

from gym-collision-avoidance.