Which repo is the latest one?

[n-sviridenko]

Hey guys,

Firstly, thanks for such a nice creature you made. It's incredible!

Secondly, I see 3 repos:

• https://github.com/honeynet/droidbot - An initial prototype/source code
• https://github.com/MobileLLM/AutoDroid - A paper around using it
• https://github.com/MobileLLM/DroidBot-GPT - A framework on top of an initial prototype

To clarify, to use that project in other projects, we need to use https://github.com/MobileLLM/DroidBot-GPT , correct? And the other 2 are deprecated.

Thanks,
Nikita

[yuanchun-li]

Hi Nikita, thanks for asking.

It is based on your purpose. If you want to use the automated GUI testing feature, then this repo (https://github.com/honeynet/droidbot) is what you need. We are working on a new project that uses LLM to improve the test coverage, but the new repo is not released yet.

If you want to build LLM-based agents for UI task automation, you can use either DroidBot-GPT or AutoDroid. The difference can be found in the papers (paper titles can be found in the README).

[n-sviridenko]

@yuanchun-li thanks for a quick reply! So droidbot is actually not a legacy project and is more up-to-date than the other two?

[n-sviridenko]

I did analysis through Claude and compared AutoDroid and DroidBot-GPT to droidbot. Leaving it here:

BRANCH: autodroid
• Core Architecture Changes:

• Added view graph building using networkx
• Implemented advanced state hashing and tracking
• Added view ID persistence mechanism
• Enhanced UI element descriptions

• State Management Improvements:

• New hash-based state string generation
• View state memory with important elements tracking
• Better handling of dynamic states (timestamps, IPs)
• Added view graph for relationship tracking

• Bug Fixes:

• Fixed empty view tree crash
• Improved IME text handling
• Better activity detection regex
• Removed problematic SDK 32+ minicap check
• Fixed scrolling behavior bugs

• UI Interaction Enhancements:

• Added smart clickability adjustment
• Improved scrolling logic with MAX_SCROLL_NUM
• Better touch event inheritance
• Enhanced text merging with safeguards

• AI Integration:

• Added hybrid LLM+graph approach
• Memory-aware prompt generation
• Support for both LMQL and standard LLM

BRANCH: droidbot-gpt
• Core Architecture Changes:

• Removed RL framework
• Simplified state representation
• Added task-based input policy
• Enhanced action description system

• State Management Improvements:

• Simplified state tracking
• Better action history logging
• Cleaner state description format

• Bug Fixes:

• Fixed IME disable check
• Improved activity detection regex
• Better event handling

• UI Interaction Enhancements:

• Improved action description generation
• Better human-readable state descriptions
• Enhanced input event handling

• AI Integration:

• Pure LLM-based decision making
• Task-oriented prompt engineering
• Streamlined LLM interaction

COMPARISON:

1. Architectural Philosophy:

• autodroid: Hybrid approach combining structured analysis with AI
• droidbot-gpt: Pure AI approach with simplified structure

2. Stability & Robustness:

• autodroid: More comprehensive error handling, better edge case coverage
• droidbot-gpt: Simpler architecture with fewer potential failure points

3. State Management:

• autodroid: Complex but more robust with graph-based tracking
• droidbot-gpt: Simpler, more straightforward but potentially less precise

4. AI Integration:

• autodroid: Sophisticated combination of graph analysis and LLM
• droidbot-gpt: Pure LLM approach with simpler integration

5. Bug Fixes:

• autodroid: More comprehensive fixes, especially for edge cases
• droidbot-gpt: Basic fixes for common issues

6. Code Complexity:

• autodroid: Higher complexity but more features
• droidbot-gpt: Lower complexity, more maintainable

7. Extensibility:

• autodroid: Better for complex automation scenarios
• droidbot-gpt: Better for simple task-based automation

8. Performance Impact:

• autodroid: Higher resource usage but more capabilities
• droidbot-gpt: Lower resource usage but more network dependent

Overall Assessment:

• autodroid appears to be a more production-ready solution with better stability and more features, but higher complexity
• droidbot-gpt offers a simpler, more focused solution that might be easier to maintain but with fewer advanced features

The choice between them would depend on specific needs:

• Choose autodroid for complex automation requiring high reliability
• Choose droidbot-gpt for simpler task-based automation where maintainability is key

Both branches represent valid approaches to the problem, but autodroid clearly invested more in stability and edge case handling, while droidbot-gpt focused on simplification and ease of use.

[yuanchun-li]

@yuanchun-li thanks for a quick reply! So droidbot is actually not a legacy project and is more up-to-date than the other two?

droidbot is not a legacy project, but I haven't updated it for quite a while. Now sure whether it will run into problems for new devices and Android versions. It should be okay for systems earlier than Android 31.