UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
A staff of researchers from UC Berkeley have launched CUA-Lite, an open platform for computer-use brokers (CUAs). The argument behind it’s infrastructural moderately than model-centric: coaching and benchmarking a CUA requires 4 items: brokers, environments, traces, and a framework to judge and prepare them and all 4 are at present fragmented throughout separate repositories with incompatible interfaces. CUA-Lite places them behind one motion house, one knowledge schema, and one command, throughout desktop, browser and cell.
Is it deployable? Yes. The stack installs with uv sync --all-extras on Python 3.12, and its light-weight sandboxes run on any Docker host with out /dev/kvm, so cloud situations, CI runners and nested containers all work.
The VM tax, and how Lite.OSWorld removes it
The most concrete contribution is Lite.OSWorld. OSWorld offers a devoted Ubuntu desktop, however it ships as a full QEMU/KVM digital machine per process, requiring nested virtualization that the majority managed infrastructure doesn’t expose. CUA-Lite reproduces the identical process suite and the identical evaluators on a GNOME desktop inside a plain Docker container.
| Task | OSWorld | Lite.OSWorld |
|---|---|---|
| Runtime | QEMU/KVM VM | Docker container |
| Host requirement | /dev/kvm, nested virt |
Any Docker host |
| Memory | 4.1 GB | 0.9 GB |
| Cold begin | 29.9 s | 23.8 s |
| Parallelism | baseline | ~4.6× extra situations |
| Task suite | OSWorld | Identical |
Fidelity is the apparent concern once you swap a VM for a container, and the staff addresses it straight: throughout 13 fashions, Lite.OSWorld scores match the OSWorld VM’s, so a rating or a coaching sign earned within the container transfers again to the actual benchmark. The similar base now carries a household of sandboxes: Lite.ScaleCUA, Lite.CUAGym and Lite.CUAWorld, the final increasing into roughly 40 purposes together with Blender, QGIS and VS Code. In whole the platform claims 30k+ verifiable duties.
One schema for knowledge, one adapter per mannequin
CUA-Lite’s second layer is LitePattern, a single supervised-learning schema shared throughout each setting, agent and process sort, shipped as plain parquet plus photographs. Ten-plus present CUA datasets have been preprocessed into it and printed free on Hugging Face, together with Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey and Multimodal-Mind2Web. Alongside these corpora sit contemporary rollout datasets generated by rolling a frontier trainer mannequin by the sandboxes, for distillation into smaller college students.
Because mannequin households anticipate totally different scaffolding, the framework ships a per-model adapter that packs a unified LitePattern into every mannequin’s personal coaching format, together with historical past collapsing so a number of steps share one ahead cross.
Eval, SFT and RL behind one command
Agents and environments meet in lite.health club: screenshots up, actions down, with one motion house per platform. 10+ brokers are inbuilt GPT, Claude, Gemini, Qwen3-VL, UI-TARS, Fara-7B, MAI-UI and others, and 15+ benchmarks are built-in, spanning grounding (ScreenSpot-Pro, OSWorld-G), desktop (OSWorld, OSWorld-2, Home windowsAgentArea, CUABench), browser (WebArena, VisualWebArena, MiniWoB, WebVoyager, Online-Mind2Web, WebGym) and cell (AndroidWorld, AndroidLab, MobileWorld, MobileGymnasium). Swapping --model-id and --env-id in scripts/rollout.py is the entire interface.
The similar loop serves coaching. For SFT, the README paperwork fine-tuning Qwen3-VL-2B-Instruct on Lite.ScaleCUA desktop trajectories, lifting imply episode return from 0.138 to 0.237 on the 332-task lite.osworld eval break up, a single reported configuration on two GPUs, not an independently reproduced outcome. For RL, rollouts scored within the setting drive GRPO updates on high of Slime, with a labored MobileGymnasium instance protecting 416 cell duties throughout 28 apps.