UC Berkeley team releases unified platform for computer-use agents
Researchers led by UC Berkeley have released CUA-Lite, an open platform designed to bring computer-use agents, environments, interaction traces, and evaluation and training tools under common interfaces. It uses one action space and one data schema across desktop, browser, and mobile settings. The software installs on Python 3.12 and its sandboxes run on Docker hosts without /dev/kvm, including cloud instances, continuous-integration runners, and nested containers.
Its main sandbox, Lite.OSWorld, reproduces the OSWorld task suite and evaluators in a GNOME desktop running inside a Docker container rather than a QEMU/KVM virtual machine. The container uses 0.9 GB of memory versus 4.1 GB for the VM and starts in 23.8 seconds versus 29.9 seconds. Across 13 models, reported scores matched those from the VM. The platform lists more than 30,000 verifiable tasks and includes additional environments for applications such as Blender, QGIS, and VS Code.
CUA-Lite also provides LiteSample, a Parquet-and-image format, with more than 20 datasets available on Hugging Face. Its single command supports evaluation, supervised fine-tuning, and reinforcement learning, with more than 10 agents and 15 benchmarks integrated. The repository currently has no explicit license, so commercial-use terms remain unclear.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.