vita-epfl / CrowdNav

[ICRA19] Crowd-aware Robot Navigation with Attention-based Deep Reinforcement Learning
MIT License
601 stars 172 forks source link

Action space question #41

Open hurong971013 opened 3 years ago

hurong971013 commented 3 years ago

I'm sorry to bother you. Excuse me, in 80 discrete action spaces, why I'm sorry to bother you. Excuse me, in 80 discrete action spaces, why should the linear velocity be exponentially valued?

ChanganVR commented 3 years ago

This is because we want the robot the take some more fine-grained actions when it gets close to the goal. Exponentially sampling leads to more low-valued velocities. We experimented with linear and expoential and found exponential worked better.

hurong971013 commented 3 years ago

This is because we want the robot the take some more fine-grained actions when it gets close to the goal. Exponentially sampling leads to more low-valued velocities. We experimented with linear and expoential and found exponential worked better.

Thank you very much. I can understand.