Berkeley RDI Opens CUA-Lite, a Shared Platform for Computer-Use Agents

UC Berkeley’s Center for Responsible Decentralized Intelligence published CUA-Lite in September 2026 as an open platform for computer-use agents. The project’s primary write-up is the RDI blog post; the companion site is cua-lite.github.io, with code at github.com/cua-lite/cua-lite and supervised corpora under the Hugging Face cua-lite org.

CUA-Lite targets a familiar bottleneck: computer-use work needs environments, labeled trajectories, and model-specific harnesses, but those pieces usually ship with incompatible interfaces. The platform standardizes three abstractions — Lite.Gym for environments, Lite.Sample for supervised data, and one harness per model shared across evaluation, supervised fine-tuning, and reinforcement learning.

What’s already plugged in

Per the RDI post, Lite.Gym exposes a Gym-style reset/step/close loop with a shared observation shape and platform action tools such as click/type on desktop and browser or tap/swipe on mobile. More than 15 benchmarks are integrated, spanning suites such as OSWorld, WebArena, and AndroidWorld. Optional VM-free Docker sandboxes (including Lite.OSWorld) aim to run large numbers of verifiable desktop tasks without /dev/kvm, while original VM-based runners remain available beside them.

Lite.Sample converts existing computer-use corpora and fresh frontier rollouts into one Hugging Face–hosted schema (parquet plus images). The post cites more than 10 datasets already converted and points to collections such as WebGym under huggingface.co/datasets/cua-lite. Model harnesses are described for 14 families so the same adapter path can drive eval rollouts and SFT exports.

Licensing caveat

CUA-Lite presents itself as an open research platform with free Hugging Face data, but public write-ups still note that the GitHub tree may not yet carry a clear SPDX license. This brief does not treat the code as Apache- or MIT-cleared for commercial redistribution until maintainers publish explicit terms.

Primary sources are the Berkeley RDI CUA-Lite blog post (September 2026), cua-lite.github.io, github.com/cua-lite/cua-lite, and the Hugging Face cua-lite dataset org.