Join us on Slack

OpenAdapt: AI-First Process Automation with Transformers

Enormous volumes of mental labor are wasted on repetitive GUI workflows.

Foundation Models (e.g. GPT-4, ACT-1) are powerful automation tools.

OpenAdapt connects Foundation Models to GUIs:

(Slides)

Welcome to OpenAdapt! This Python library implements AI-First Process Automation with the power of Transformers by:

• Recording screenshots and associated user input
• Aggregating and visualizing user input and recordings for development
• Converting screenshots and user input into tokenized format
• Generating synthetic input via transformer model completions
• Replaying synthetic input to complete tasks

The goal is similar to that of Robotic Process Automation, except that we use transformers instead of conventional RPA tools.

The direction is adjacent to Adept.ai, with some key differences:

1. OpenAdapt is model agnostic;
2. OpenAdapt generates prompts automatically (auto-prompted, not user-prompted);
3. OpenAdapt works with all types of desktop GUIs, including virtualized (e.g. Citrix) and web
4. OpenAdapt is open source! (license TBD, please see #246)

Install

Recommended: Install with Poetry:

git clone https://github.com/MLDSAI/OpenAdapt.git
cd OpenAdapt
pip install poetry
poetry install
poetry shell
alembic upgrade head
pytest

Manual:

git clone https://github.com/MLDSAI/OpenAdapt.git
cd OpenAdapt
python3.10 -m venv .venv
source .venv/bin/activate
pip install wheel
pip install -r requirements.txt
pip install -e .
python -m spacy download en_core_web_trf
alembic upgrade head
pytest

Permissions

See how to set up system permissions on macOS here.

Run

Record

Create a new recording by running the following command:

python -m openadapt.record "testing out openadapt"

Wait until all three event writers have started:

| INFO     | __mp_main__:write_events:230 - event_type='screen' starting
| INFO     | __mp_main__:write_events:230 - event_type='action' starting
| INFO     | __mp_main__:write_events:230 - event_type='window' starting

Type a few words into the terminal and move your mouse around the screen to generate some events, then stop the recording by pressing CTRL+C.

Current limitations:

• recording should be short (i.e. under a minute), as they are somewhat memory intensive, and there is currently an open issue describing a possible memory leak
• the only touchpad and trackpad gestures currently supported are pointing the cursor and left or right clicking, as described in this open issue

Visualize

Visualize the latest recording you created by running the following command:

python -m openadapt.visualize

This will open your browser. It will look something like this:

Playback

You can play back the recording using the following command:

python -m openadapt.replay NaiveReplayStrategy

More ReplayStrategies coming soon! (see Contributing).

Contributing

Problem Statement

Our goal is to automate the task described and demonstrated in a Recording. That is, given a new Screenshot, we want to generate the appropriate ActionEvent(s) based on the previously recorded ActionEvents in order to accomplish the task specified in the Recording.task_description, while accounting for differences in screen resolution, window size, application behavior, etc.

If it's not clear what ActionEvent is appropriate for the given Screenshot, (e.g. if the GUI application is behaving in a way we haven't seen before), we can ask the user to take over temporarily to demonstrate the appropriate course of action.

Dataset

The dataset consists of the following entities:

1. Recording: Contains information about the screen dimensions, platform, and other metadata.
2. ActionEvent: Represents a user action event such as a mouse click or key press. Each ActionEvent has an associated Screenshot taken immediately before the event occurred. ActionEvents are aggregated to remove unnecessary events (see visualize.)
3. Screenshot: Contains the PNG data of a screenshot taken during the recording.
4. WindowEvent: Represents a window event such as a change in window title, position, or size.

You can assume that you have access to the following functions:

• create_recording("doing taxes"): Creates a recording.
• get_latest_recording(): Gets the latest recording.
• get_events(recording): Returns a list of ActionEvent objects for the given recording.

Instructions

Join us on Slack. Then:

1. Fork this repository and clone it to your local machine.
2. Get OpenAdapt up and running by following the instructions under Setup.
3. Look through the list of open issues at https://github.com/MLDSAI/OpenAdapt/issues and once you find one you would like to address, indicate your interest with a comment.
4. Implement a solution to the issue you selected. Write unit tests for your implementation.
5. Submit a Pull Request (PR) to this repository. Note: submitting a PR before your implementation is complete (e.g. with high level documentation and/or implementation stubs) is encouraged, as it provides us with the opportunity to provide early feedback and iterate on the approach.

Evaluation Criteria

Your submission will be evaluated based on the following criteria:

1. Functionality : Your implementation should correctly generate the new ActionEvent objects that can be replayed in order to accomplish the task in the original recording.

2. Code Quality : Your code should be well-structured, clean, and easy to understand.

3. Scalability : Your solution should be efficient and scale well with large datasets.

4. Testing : Your tests should cover various edge cases and scenarios to ensure the correctness of your implementation.

Submission

1. Commit your changes to your forked repository.

2. Create a pull request to the original repository with your changes.

3. In your pull request, include a brief summary of your approach, any assumptions you made, and how you integrated external libraries.

4. Bonus: interacting with ChatGPT and/or other language transformer models in order to generate code and/or evaluate design decisions is encouraged. If you choose to do so, please include the full transcript.

We're hiring!

If you're interested in getting paid for your work, please mention it in your Pull Request.

Troubleshooting

MacOS: if you encounter system alert messages or find issues when making and replaying recordings, make sure to set up permissions accordingly.

In summary (from https://stackoverflow.com/a/69673312):

1. Settings -> Security & Privacy
2. Click on the Privacy tab
3. Scroll and click on the Accessibility Row
4. Click +
5. Navigate to /System/Applications/Utilities/ (or wherever Terminal.app is installed)
6. Click okay.

Developing

Generate migration (after editing a model)

alembic revision --autogenerate -m "<msg>"

Submitting an Issue

Please submit any issues to https://github.com/MLDSAI/openadapt/issues with the following information:

• Problem description (please include any relevant console output and/or screenshots)
• Steps to reproduce (please help others to help you!)