Training data for BEST_WALK_ONNX_2.onnx

Hello, I am experimenting with BEST_WALK_ONNX_2.onnx, see [example](https://github.com/Juliaj/ros2_control_demos/tree/onnx_demo_open_duck_mini/example_18). I'm running Mini Duck in Mujoco simulation.

I've observed some instability in the inference scenario, in which Duck falls down immediately after receiving a command to move forward. I'm investigating whether this is due to observation distribution shift in my env. Would you be able to provide some training data (observations plus actions) for BEST_WALK_ONNX_2.onnx ?

It appears that the model expects 3 previous actions (last_action, last_last_action, last_last_last_action). How should these be initialized with the first command, meaning when Duck stands still? Tips and suggestions are appreciated.