search

bradyz / 2020_CARLA_challenge

"Learning by Cheating" (CoRL 2019) submission for the 2020 CARLA Challenge
181 stars 49 forks source link

Distribution of navigational commands in the full dataset #9

Closed ap229997 closed 4 years ago

ap229997 commented 4 years ago

The distribution of the navigational commands in the full dataset (# images for each command) is severely skewed towards command '4'. Is this due to the autopilot or the 76 routes provided in leaderboard/data/routes_*.xml? carla_9

bradyz commented 4 years ago

This is due to the preset routes. Most of the time is simply lane following (which I have merged with straight)

ap229997 commented 4 years ago

Training the MapAgent on this data would probably result in a very biased agent that would perform poorly at turns and lane changes?

bradyz commented 4 years ago

in our experiments, we've seen the map agent to do quite well due to the nice representation of the map

the image agent, if trained naively, fails in the ways you mentioned to do better across all navigational commands, we use the all-branches supervision from LBC, which acts as a sort of nice sampling

note: in this repo, we don't condition on discrete actions, but actually render a sort of "command heatmap" that you can see in the readme. I found these to work better in the CARLA challenge

ap229997 commented 4 years ago

Thanks for the clarification.