- Published on
UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-U...
- Authors
- Name
- DeepHorus
UC Berkeley researchers have unveiled CUA-Lite, a new open-source platform designed to streamline the development and testing of computer-use agents (CUAs). The release addresses a core problem in the field: the fragmentation of essential componentsâagents, environments, data traces, and training frameworksâacross incompatible repositories. CUA-Lite unifies these elements under a single action space, data schema, and command structure for desktop, browser, and mobile environments, aiming to accelerate research and development.
The platform's foundational contribution is Lite.OSWorld, which brings the comprehensive Ubuntu desktop evaluation suite to lightweight Docker containers. This eliminates the need for the full QEMU/KVM virtual machines required by the original OSWorld, which demand nested virtualization often unavailable in cloud or CI environments. Researchers report that scores on Lite.OSWorld match the original VM benchmarks across 13 different models, ensuring that training signals are transferable.
The technical efficiency of this containerized approach is significant. Lite.OSWorld occupies approximately 0.9 GB, compared to the 4.1 GB of the VM, allowing for roughly 4.6 times more parallel desktops on the same hardware. This dramatically reduces the infrastructure cost and complexity for large-scale agent training and evaluation. The base container also supports a family of expanded sandboxes, including Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld.
A second key layer is LiteSample, a standardized data schema for supervised learning. It packages trajectories as plain Parquet files with images, ensuring consistency across all environments and tasks. More than ten existing CUA datasets, such as Aguvis and ScaleCUA, have been preprocessed into this format and made freely available on Hugging Face. Fresh rollout datasets from frontier models are also included to facilitate knowledge distillation.
The platform integrates this data pipeline with its core evaluation and training loop. The lite.gym module provides a unified interface where agents process screenshots and return actions, with 10+ built-in agents and over 15 integrated benchmarks. The same command structure supports both supervised fine-tuning (SFT) and reinforcement learning (RL). A documented example shows SFT improving Qwen3-VL-2B's performance on a desktop task suite from 0.138 to 0.237 mean episode return.
In total, CUA-Lite claims to offer over 30,000 verifiable tasks within a deployable stack that works on any Docker host. By consolidating the entire workflow from data to deployment, the researchers aim to lower barriers to entry and foster more reproducible, scalable innovation in building agents that can autonomously use computers. The open release is expected to provide a common foundation for both academic and industry research.
