BrowserGym

BrowserGym copied to clipboard

some library can be particularly verbose if we set debug log. Let's make sure exp_args.log_level only affects bgym's logs

recursix

I think we tried to adjust in the past, but it still there. The screenshots of miniwob are still contained within a large gray rectangles. It would be great if...

recursix

# What we want The current chat API is not super modular. Ideally we would like to be able to switch between different chat implementations, which can be parameterized via...

gasse

Currently BrowserGym's `task.cheat()` functions use a list of messages to interact with the chat: ```python def cheat(self, page: playwright.sync_api.Page, chat_messages: list[str]): ``` It should take the same arguments as `execute_python_code`:...

gasse

Right now, it allows passing an id during `__init__`: https://github.com/ServiceNow/BrowserGym/blob/04bd94487be21c399a7c78cbb4df505e30d2fbb9/webarena/src/browsergym/webarena/task.py#L23-L85 We could add another parameter `task_configs` which is of type `List[Union[dict, str]]`, for example: ``` GenericWebArenaTask(task_configs=[{...}, {...}], ...) ``` where...

xhluca

Cross Domain Iframes missing from DOM and Accessibility Tree text

1

Example: https://www.nexhealth.com/careers/open-positions?gh_jid=5256431004 See https://issues.chromium.org/issues/333981074. I also tried calling playwright's page.frames to get all the frames on a page with 2 levels of frame nesting, but a grandchild frame was missing,...

cortezpage

------------------- ## 4x : Exception uncaught by agent or environment in task .KeyboardInterrupt:Early termination? * visualwebarena.317 seed: 31 * visualwebarena.317 seed: 31 * visualwebarena.319 seed: 0 * visualwebarena.319 seed: 0...

recursix

task hanging in VWA with vision agent

4

It's weird that the non-vision agent had much less error than the vision based agent. ```bash INFO:root:All jobs are finished. Calling agent_args.close() on all agents... INFO:root:Experiment finished. Searching experiments directories.:...

recursix

Task hanging in webarena

2

Logs were not in debug mode: ``` 2024-10-24 05:41:36,037 - 766755 - browsergym.experiments.loop - INFO - Running experiment GenericAgent-gpt-4o-mini-2024-07-18_on_visualwebarena.241_16 in: /home/toolkit/agentlab_results/2024-10-24_02-15-43_GenericAgent-gpt-4o-mini-2024-07-18_on_visualwebarena/2024-10-24_02-15-44_GenericAgent-gpt-4o-mini-2024-07-18_on_visualwebarena.241_16 2024-10-24 05:41:56,460 - 766755 - root - WARNING -...

recursix

Opening a PDF results in a hang on x86_64 (Linux or MacOS) in headless mode. Seems to be working fine on arm64 (MacOS M1) either in headless or headfull mode.

gasse