UC Berkeley Researchers Release CUA-Lite for Computer-Use Agents

Researchers at the University of California, Berkeley, have released CUA-Lite, an open-source platform designed to streamline the development and evaluation of computer-use agents (CUAs). The platform addresses a fragmentation issue in the CUA ecosystem, where agents, environments, data schemas, and training/evaluation frameworks are typically scattered across disparate repositories with incompatible interfaces. CUA-Lite consolidates these four essential components into a single, unified system accessible through a singular action space, data schema, and command, supporting desktop, browser, and mobile environments. This unified approach aims to simplify the infrastructure required for training and benchmarking CUAs.

The platform's core contribution is Lite.OSWorld, which offers a lightweight alternative to OSWorld's resource-intensive virtual machine approach. OSWorld typically utilizes a full QEMU/KVM virtual machine for each task, necessitating nested virtualization that is often unavailable in managed cloud infrastructure. In contrast, CUA-Lite replicates OSWorld's task suite and evaluators within a GNOME desktop environment running inside a standard Docker container. This Docker-based implementation significantly reduces resource requirements, with Lite.OSWorld consuming 0.9 GB of memory compared to OSWorld's 4.1 GB. Furthermore, it eliminates the need for /dev/kvm and nested virtualization, making it compatible with any Docker host, including cloud instances, CI runners, and nested containers. The cold start time for Lite.OSWorld is also improved, at 23.8 seconds, versus OSWorld's 29.9 seconds. Benchmarking indicates that CUA-Lite can run approximately 4.6 times more instances in parallel compared to the baseline OSWorld VM setup.

Crucially, CUA-Lite maintains fidelity with the original OSWorld benchmarks. The research team reports that across 13 different AI models, Lite.OSWorld achieves scores consistent with those obtained using the OSWorld VM. This ensures that training signals and performance evaluations conducted within the containerized environment are transferable to the established VM-based benchmarks. The platform supports a family of sandboxes built on this common base, including Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld. Lite.CUAWorld, in particular, expands the range of supported applications to approximately 40, encompassing software like Blender, QGIS, and VS Code. In total, the CUA-Lite platform claims to support over 30,000 verifiable tasks.

Beyond the sandboxes, CUA-Lite introduces LiteSample, a unified schema for supervised learning data that is consistent across all supported AI models. This single schema simplifies data management and integration, eliminating the need for model-specific data adapters. The stack is designed for straightforward deployment, installable on Python 3.12 using `uv sync --all-extras`. The lightweight nature of its sandboxes allows them to run on any Docker host, removing the "VM tax" and making advanced CUA development and testing more accessible and cost-effective.