OSWorld/mm_agents

Agent

Prompt-based Agents

Supported Models

We currently support the following models as the foundational models for the agents:

• GPT-3.5 (gpt-3.5-turbo-16k, ...)
• GPT-4 (gpt-4-0125-preview, gpt-4-1106-preview, ...)
• GPT-4V (gpt-4-vision-preview, ...)
• Gemini-Pro
• Gemini-Pro-Vision
• Claude-3, 2 (claude-3-haiku-2024030, claude-3-sonnet-2024022, ...)
• ...

And those from the open-source community:

• Mixtral 8x7B
• QWEN, QWEN-VL
• CogAgent
• ...

In the future, we will integrate and support more foundational models to enhance digital agents, so stay tuned.

How to use

from mm_agents.agent import PromptAgent

agent = PromptAgent(
    model="gpt-4-vision-preview",
    observation_type="screenshot",
)
agent.reset()
# say we have an instruction and observation
instruction = "Please help me to find the nearest restaurant."
obs = {"screenshot": open("path/to/observation.jpg", 'rb').read()}
response, actions = agent.predict(
    instruction,
    obs
)

Observation Space and Action Space

We currently support the following observation spaces:

• a11y_tree: the accessibility tree of the current screen
• screenshot: a screenshot of the current screen
• screenshot_a11y_tree: a screenshot of the current screen with the accessibility tree overlay
• som: the set-of-mark trick on the current screen, with table metadata included.

And the following action spaces:

• pyautogui: valid Python code with pyautogui code valid
• computer_13: a set of enumerated actions designed by us

To feed an observation into the agent, you have to maintain the obs variable as a dict with the corresponding information:

# continue from the previous code snippet
obs = {
    "screenshot": open("path/to/observation.jpg", 'rb').read(),
    "a11y_tree": ""  # [a11y_tree data]
}
response, actions = agent.predict(
    instruction,
    obs
)

Efficient Agents, Q* Agents, and more

Stay tuned for more updates.