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
166 stars 29 forks source link

Question about LBC dataset #31

Closed zqh0253 closed 2 years ago

zqh0253 commented 2 years ago

Hi, dian! Thanks for releasing this wonderful work. Currently I have some questions about the setup of LBC dataset (for training the cheating agent). In your original paper, the expert of the privileged agent is autopilot policy. So I think the agent used to collect dataset should be autopilot agent. But according to the README of LBC in this repo, the dataset follows RAILS' setup procedure and seems to be built by (1) collecting dataset with random agents (2) Q learning (3) collecting dataset with Q-agent. This procedure is different as specified in the original paper.

I'm wondering if I've misunderstood something. If yes, please kindly correct me. If not, does these two types of dataset bring performance gap on the privileged agent?

dotchen commented 2 years ago

Hi,

Thank you for your interest in our project.

To answer your question shortly: the Q-collector here is the expert. The one that comes with the official CARLA repo in 0.9.10 does not handle pedestrians.

zqh0253 commented 2 years ago

Thanks for your answer!