search

dotchen / LearningByCheating

(CoRL 2019) Driving in CARLA using waypoint prediction and two-stage imitation learning
MIT License
308 stars 102 forks source link

Number of pedestrians in dense scenario of No-Crash benchmark Town1/Town2 #1

Closed marintoro closed 4 years ago

marintoro commented 5 years ago

Hello,

Really nice code and re-implementation of the CARLA benchmark! I found something weird, in the benchmark/init.py where the scenarios are defined, it's written that the NoCrashTown01-v5 scenario got 250 pedestrians while the NoCrashTown02-v5 has only 150. However, I launched both scenario and it seems like there is much much more pedestrians spawned and crossing roads in NoCrashTown02-v5. Is that a bug? Or maybe I didn't understand something there...

dotchen commented 4 years ago

Thank you for your interest in our project. You could take a look at the logic for spawning pedestrians here: https://github.com/dianchen96/LearningByCheating/blob/release-0.9.6/bird_view/utils/carla_utils.py#L422. Note that Town2 is smaller than Town1 and has much smaller walkable region defined in the navmesh. This results in a higher pedestrian density at the walkable regions. Please let us know if you still believe there is a bug.

marintoro commented 4 years ago

Thanks for this fast reply. I think the Town2 is smaller than Town1 and has much smaller walkable region totally answer my question! As a results my agent is way worse in Town2 dense scenario than Town1 dense scenario even if I actually trained it on Town2.