search

jchengai / planTF

[ICRA'2024] Rethinking Imitation-based Planner for Autonomous Driving
https://jchengai.github.io/planTF/
195 stars 12 forks source link

Questions about ego history #17

Open zhengyang-ponyai opened 5 months ago

zhengyang-ponyai commented 5 months ago

Really insightful work! I want to know if you apply SDE to history information, will it force network not learn history shortcut? Are you think ego history is completely useless or it just can't learn well by model?

jchengai commented 5 months ago

Hi @zhengyang-ponyai, history dropout will also work (as the case in ChauffeurNet). And here are my thoughts:

  1. I think ego history is not completely useless. For example, in integrated planning & prediction, it might be more appropriate for other agents to have access to ego's history. And model trained with history has clearly better prediction metrics (ADE, FDE). This is because history contains higher dimension kinematics such as jerk, yaw rate, et al. So it is ok to use history if you can prevent the model relying too much on history info.

  2. For planning task, the current state is enough to do planning since we have a goal and route. And humans do not need to remember their histories to drive.

  3. It is much simpler to augment the state-based model. Otherwise, you have to generate a new history for ego after perturbation.