search

facebookresearch / deep_bisim4control

Learning Invariant Representations for Reinforcement Learning without Reconstruction
Other
143 stars 37 forks source link

Questions on reproducing CARLA #14

Open ttaa9 opened 3 years ago

ttaa9 commented 3 years ago

Hi, thanks for the great paper and comprehensive code release. I had some questions about reproducing the CARLA results.

  1. Are the correct hyper-parameters (to reproduce the paper results) in run_local_carla096.sh? For instance, I notice that encoder_stride=2 there, but the paper appendix appears to state it is 1.
  2. I noticed that the README instructions utilize CARLA_0.9.8 (including adding the 0.9.8 Python API to the path, meaning we will use the carla_env.py for 0.9.8), but actually runs the CARLA_0.9.6 CarlaUE4.sh script. Is this correct?
  3. I noticed in #8 that the ensemble model was necessary to perform well. Is that the case here too?

Thanks!

YunpengZhai commented 1 year ago

Hi, @ttaa9 , have you solved your problem? I am blocked with the same problem. Thanks!