search

vita-epfl / CrowdNav

[ICRA19] Crowd-aware Robot Navigation with Attention-based Deep Reinforcement Learning
MIT License
560 stars 166 forks source link

Algorithm Problem #59

Open ShelyH opened 1 year ago

ShelyH commented 1 year ago

Hi, sorry to bother you again! When I use the lstm_rl algorithm, the navigation paths show oscillation and freezing behavior, why is this? Could you give me some suggestions?

ChanganVR commented 1 year ago

hi @2529342549 these networks are trained end-to-end and it's hard to interpret individual policy's behavior. Although I don't really know, this behavior could actually be indicating what could be done for improvement :)

ShelyH commented 1 year ago

Hi, sorry to bother you again! The hidden state of this part of the predict in lstm-rl is assigned zero each time, not the hidden state in the previous state. What is the role of LSTM in this case?

ChanganVR commented 1 year ago

@2529342549 I was not aware of that at all. Could be a bug. Have you debugged it?