Guide to deploy it on Android devices

[dubeyabhijeet]

Hi ,

I want to use it on real devices not on emulator.
Can you help me guide how to do it?

Regards,
Abhijeet

[kenjitoyama]

Hello @dubeyabhijeet! We do not officially support running AndroidEnv on real devices at the moment.
To do that, you'll need a new Simulator to grab screenshots and sending actions. If you don't care about performance that much, it's relatively straightforward to write one that uses adb screencap (https://developer.android.com/studio/command-line/adb#screencap) and adb input tap. Keep in mind that this solution will probably achieve only a handful of steps per second.

Cheers,

Daniel

[dubeyabhijeet]

Understood ! Thanks a lot....

Right now you are controlling it from PC to emulator. If i used real devices then performance will be trade off [PC to real devices].

What about deploy it through apk in phone? Like i use app traversal and on each screen use android_env to decide on actions to take? Like ndroid_env in wrapped in traversal apk and deployed on phone.

[kenjitoyama]

Yes, it's possible but at the moment there's no good way to do so. There are no APIs to fetch screenshots and execute actions from within Android, and that would also probably require root access, which is a security issue.

[dubeyabhijeet]

Got it ! Understood. Thanks a lot Daniel :)

[dubeyabhijeet]

Hi Daniel, What i see we need to add logs in apps source code. I have two queries:

1. Not possible with 3rd party apps without source code?
2. How deepmind trained agents for youtube? Have source code or youtube has those logs needed for rewards?

[kenjitoyama]

Hello @dubeyabhijeet!

Not possible with 3rd party apps without source code?

You can run anything you want in the emulator, but rewards are only available if something is shown in the log stream and captured by a regex.

How deepmind trained agents for youtube? Have source code or youtube has those logs needed for rewards?

We haven't trained explicitly for YouTube, it was just an example. However, we did train agents on apps for which we don't have the source code such as the Clock or System Settings by reusing accessibility events. Please see https://github.com/deepmind/android_env/blob/main/docs/example_tasks.md#accessibility-forwarder. At one point, we also extracted rewards from view hierarchies, but they're much slower to get (it needs a slow adb dumpsys call + parsing) than simple Log.i() calls. We also experimented with OCR, but though that works in synthetic text in static apps, it doesn't really work well for things like games or apps with unusal fonts and/or graphics. In theory, any event or message in Android could be used as a source of rewards, but we relied on logs because they're the easiest to implement (from a developer and task designer's point of views) and reliable.

Cheers,

Daniel