Uc berkeley launches cua-lite: a unified platform for training and evaluating computer-use agents

Researchers from UC Berkeley have introduced CUA-Lite, an open-source platform designed to streamline the development and evaluation of computer-use agents (CUAs). The platform addresses the fragmentation of existing tools by integrating agents, environments, traces, and evaluation frameworks into a unified system. CUA-Lite supports multiple platforms, including desktop, browser, and mobile, through a standardized action space, data schema, and command structure. A key component of CUA-Lite is Lite.OSWorld, which provides a faithful Ubuntu desktop experience within a Docker container, eliminating the need for nested virtualization. The platform also includes Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld, supporting over 40 applications and 30,000+ verifiable tasks. Additionally, LiteSample offers a shared supervised-learning schema across environments, agents, and tasks, with preprocessed datasets available on Hugging Face. The platform supports both supervised fine-tuning and reinforcement learning, with examples demonstrating improvements in task performance. CUA-Lite aims to facilitate collaboration and innovation in the field of computer-use agents by providing a comprehensive and accessible framework.