Questions about unified action space and Benchmrk action categories ？

[zhshj0110]

Thanks for your excellent work.

I found that the performance issues on GUI-Odyeesy and Android-Control were not reproducible.

• Could you provide the best parameter Settings for offline evaluation #9
• How can I reproduce your result in AndroidWorld? Could you provide some scripts for users to reproduce it easily? #11
• I can't reproduce the test results on Odyssey and Android_Control #18

During training and evaluation, did you customize prompts and action spaces for each dataset ? For example, GUI-Odyeesy does not have Launch(app=app_name), but Android-Control has a similar action?

By using the same action space as the benchmark, the model reduces the number of unexpected actions predicted, and performance improves.