WebHarbor
A test and RL training suite for GUI agents comprising 153 Dockerized executable website environments. It features high visual fidelity and functional depth, and supports multi-step tasks and tasks requiring login.
Timeline (last 90 days)
WebHarbor is a test and training suite comprising 153 executable, Dockerized website environments, proposed by the aiming-lab research team.
The WebHarbor paper and code have been released in the GitHub repository aiming-lab/WebHarbor.
WebHarbor is built through a task-driven human-AI collaborative process, in which coding agents implement functional changes and human contributors review visual fidelity, functional correctness, and task validity.
The WebHarbor suite contains a total of 3,518 tasks, many of which involve multi-step flows and post-login workflows.
The research team used RL to train a Qwen 4B-level GUI agent in WebHarbor.
The Qwen 4B-level GUI agent trained with RL in WebHarbor can generalize to unseen tasks and performs comparably to leading open-source GUI web agents on four benchmarks.
153 environments are still limited in scale relative to the real internet, and environment evolution depends on sustained investment in manual review.
All Facts (7)
The WebHarbor suite contains a total of 3,518 tasks, many of which involve multi-step flows and post-login workflows.
50%UnverifiedThe Qwen 4B-level GUI agent trained with RL in WebHarbor can generalize to unseen tasks and performs comparably to leading open-source GUI web agents on four benchmarks.
50%Unverified153 environments are still limited in scale relative to the real internet, and environment evolution depends on sustained investment in manual review.
50%UnverifiedWebHarbor is built through a task-driven human-AI collaborative process, in which coding agents implement functional changes and human contributors review visual fidelity, functional correctness, and task validity.
50%UnverifiedWebHarbor is a test and training suite comprising 153 executable, Dockerized website environments, proposed by the aiming-lab research team.
50%UnverifiedThe WebHarbor paper and code have been released in the GitHub repository aiming-lab/WebHarbor.
50%UnverifiedThe research team used RL to train a Qwen 4B-level GUI agent in WebHarbor.
50%