[Question] sim2real question

[jh-s5974]

I'm writing a sim2real code

Since I use 50 obs data, I input the values ​​corresponding to each obs data in order to the onnx neural network

Then, I input the 12 action values ​​output from onnx to the robot's joints

At this time, the joint input value is input as action*action_scale value

And, the action value without multiplying the action scale is used as the previous action value and used as the input data for the new onnx file

I structured the code like this, but when I ran the code, the data generated as the joint input value gradually increased and eventually diverged to an infinite value

Why is this happening?

Just in case, I tried using the action*action_scale value as the previous action value and it worked the same way.

I used the learning created by referring to the existing bipedal walking code

I look forward to your reply.

Thank you.

[RandomOakForest]

Thank you for posting this question. Should we assume the network has already been trained? it seems there is some observation data processing missing, or the network needs more rounds of training, or both. If you could elaborate on the task at hand, and which are the pieces you are trying to put together may help us provide better feedback. I will move this post to our Discussions section for follow up. Thank you for your interest in Isaac Lab.

[jh-s5974]

Learning is complete and the robot walks well.

And the observation values ​​are entered well, and the first values ​​output as action values ​​are confirmed normally.

However, when moving to the next step, when the joint position values ​​are updated using the action * action scale values, the values ​​start to increase and diverge.