Question ML training brain (ONNX) causes kart racer to start in random location

After training my kart racer using

Code (csharp):

1. mlagents-learn --run-id=test1

I put the newly generated ONNX file in the "model" field of the Behavior Parameters component on the Kart Racer ML Agent car. The car runs my custom track much faster now, but the starting position of the car is some random place on the track. The intended outcome was that the car will start where I place it, which is on the start/finish line.

Is anyone familiar with this behavior? Any ideas how I can correct this?

 

the brain only makes decisions, the position is being controlled by a script

 

Hmmm, ok. I guess something else is changing the start location. If I am training the kart racer, should I be trying to load the original Neural Network model into the mlagents-learn command? Or let the mlagents-learn create a completely new model? I should note that I have created a whole different track for it to learn on.