UC Berkeley Releases CUA-Lite for Computer-Use Agents
UC Berkeley researchers release CUA-Lite, an open platform unifying sandboxes, data, evaluation, and RL for computer-use agents.
Stock photo for illustration only, not from the actual event
- UC Berkeley researchers launch CUA-Lite for computer-use agents
- Installs via Python 3.12 and runs on standard Docker without /dev/kvm
- Integrates over 10 datasets and 30,000 verifiable tasks
Researchers at UC Berkeley have developed and released CUA-Lite, a novel open-source platform designed to unify sandboxes, training data, evaluation frameworks, and reinforcement learning for computer-use agents. The stack installs effortlessly using uv sync --all-extras on Python 3.12 and operates on any standard Docker host without requiring /dev/kvm, enabling seamless deployment across cloud instances and nested containers.
A concrete contribution of the project is Lite.OSWorld. While traditional OSWorld ships as a full QEMU/KVM virtual machine per task requiring nested virtualization, CUA-Lite successfully reproduces the task suite and evaluators on a GNOME desktop inside a plain Docker container. Benchmark results across 13 models confirm that container scores match the original OSWorld VM scores.
Stock photo for illustration only, not from the actual event
The platform houses a family of sandboxes including Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld, expanding into roughly 40 applications such as Blender, QGIS, and VS Code, totaling over 30,000 verifiable tasks. Furthermore, the second layer, LiteSample, provides a unified supervised-learning schema across all environments, shipped as plain parquet files with images and published freely on Hugging Face.
The release of CUA-Lite addresses a major infrastructure bottleneck for AI agent developers. Traditional full virtual machine setups impose high computational overhead and deployment hurdles, whereas lightweight Docker containers significantly lower the barrier for researchers to scale experiments and train models efficiently.
Connecting agents and environments, lite.gym handles screenshots and actions with unified action spaces. It integrates over 10 built-in agents such as GPT, Claude, Gemini, Qwen3-VL, UI-TARS, and Fara-7B, alongside more than 15 benchmarks spanning desktop, browser, and mobile environments. Developers can switch models and environments simply by modifying script parameters.
For training workflows, the framework documents fine-tuning Qwen3-VL-2B-Instruct on Lite.ScaleCUA desktop trajectories, which raised mean episode return from 0.138 to 0.237 on the 332-task evaluation split using two GPUs. Reinforcement learning rollouts drive GRPO updates on top of Slime, demonstrated via a MobileGym example covering 416 mobile tasks across 28 applications.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment