search

dotchen / WorldOnRails

(ICCV 2021, Oral) RL and distillation in CARLA using a factorized world model
https://dotchen.github.io/world_on_rails/
MIT License
167 stars 29 forks source link

about data collecting in data_phase0 #46

Open Watson52 opened 2 years ago

Watson52 commented 2 years ago

Hi Chen. Thanks for a lot of open source works and I am trying to follow them. I jump from the data collecting about LAV as you said there are similarities between them. I found that the random collector return a random control including throttle and steer. I feel confused how could the agent arrived the goal under the random control? Does WOR and LAV use the same agent? How does it chose random routes?

dotchen commented 2 years ago

Hi,

Thanks a lot for your interest in our projects. The distinction between LAV and Rails (WOR) in data collection is: LAV uses an expert; Rails first use a completely random data to train a vehicle kinematics mode; it then uses this model to build a crude agent to collect the main traces; the main traces replay and we use model-based RL offline to distill an agent.

Let me know if you have further questions.

Watson52 commented 2 years ago

Thanks for your reply, it make me clear about the WOR. I think I should relize how the leaderboard and scenario runner work first. I will ask you if I meet problem. Thanks again.

dotchen commented 2 years ago

Sure thing. Feel free to also email me at dchen@cs.utexas.edu for more questions.