I chanced some problems about your code

[liheng12345]

I had successfully excute your code in the ros platform, but the code dose not work as well as the paper. when I train the agent 1200 times , it also show that the PID controller preforms main factors. and I have read your code ,but the desiged network seem not to be 2D convenlution layers. just use the MLP layers.

if you have time ,can you answer my questions, thanks~~~~

[akmandor]

@liheng12345 Although they present like 2D image in Figure 2 in their paper, actually they propose 1D CNN (by laser scan data 1x512 where channel is 3 for the consecutive time steps). However, you are right, I also could not find any implementation detail with respect to their proposed 1D CNN.

@xie9187 Could you share the necessary details (to replicate results in your paper) of the 1D CNN especially filter(kernel) sizes and strides at each layer?