Experiment Setup of Snake Locomotion #4
Description
Dear author, I tried to train SAC on the Snake Locomotion environment, but it seems it is difficult to train and converge as Fig. 6 shows on the paper.
I set the episode steps as 4000 and trained the algorithm for 10^7 steps. All other settings are kept default. I have a few questions and hope the authors can help me. Thank you for your time and insights!
- the action space is 16 for 8 actuators. When I opened the expert's trajectory provided here, I found the second actions for each actuator are 0. So do I need to set the second action produced by RL as 0?
- Do I need to change other settings in SAC to make it work as well as the paper shows?