Online Reinforcement Learning Agent for Web Navigation

A new research paper introduces OpAgent, an autonomous web agent that uses online reinforcement learning to navigate real-world websites. Unlike conventional methods relying on static datasets and supervised fine-tuning, OpAgent learns through direct iterative interactions with unconstrained web environments. The approach features hierarchical multi-task fine-tuning with datasets categorized into Planning, Acting, and Grounding primitives, establishing a Vision-Language Model with strong instruction-following capabilities. This addresses distributional shifts caused by offline trajectories failing to capture stochastic state transitions and real-time feedback.