motion-planner-reinforcement-learning
motion-planner-reinforcement-learning copied to clipboard
m5823779
Did you manage to solve the action saturation issue?
Hi, I see in this code you are outputting both linear and angular velocity and would like to know how you got around the action saturation problem?
I got the same problem. When outputting both linear and angular velocity, it got 1 or 0 after some gradient decent. I have checked the code and tested it using Pendulum-v0 environment. But I don't know how to fix it.
I got the same problem. When outputting both linear and angular velocity, it got 1 or 0 after some gradient decent. I have checked the code and tested it using Pendulum-v0 environment. But I don't know how to fix it.
Hello, I also meet the problem of the action saturation, I noticed that the author used a batch normalization after each layer in actor network. I'm wondering if it is the solution to fix the problem and did you fix it now?
I got the same problem. When outputting both linear and angular velocity, it got 1 or 0 after some gradient decent. I have checked the code and tested it using Pendulum-v0 environment. But I don't know how to fix it.
Hello, I also meet the problem of the action saturation, I noticed that the author used a batch normalization after each layer in actor network. I'm wondering if it is the solution to fix the problem and did you fix it now?
Hello, I have the same problem, do you have a solution? I would be grateful if you could help me!
I am using SAC and meet the same issue too. I changed many hype-parameters (lr, batch size, initial memory, state format, action range) and the issues have been fixed. I am not sure that which factor is the key, but I am pretty sure that is not batch norm, because I not using batch norm in model design.
