Berkeley RDI releases CUA-Lite, an open platform for computer-use agents

Berkeley RDI 发布开源平台 CUA-Lite，面向计算机使用智能体

Berkeley RDI open-sourced CUA-Lite, a platform that unifies environments, training data, and model interfaces for computer-use agents. It has three standardized parts: Lite.Gym wraps 15+ benchmarks and 30k+ verifiable tasks behind one API, Lite.Sample converts 10+ datasets into a single format, and model harnesses let 14 model families share the same eval, SFT, and RL pipeline. The standout is a VM-free Docker sandbox that runs OSWorld tasks without hardware virtualization, cutting costs. Code and datasets are public on GitHub and Hugging Face.

Why it matters: Berkeley RDI ships an open platform that unifies environments, data, and model interfaces for computer-use agents — directly tackling the fragmentation pain point. 15+ benchmarks, 10+ datasets, 30k+ verifiable tasks, plus a Docker sandbox that drops the hardware virtualization...