Google’s ARTEMIS Brings Natural-Language AI Automation to Android

Summary

Google’s ARTEMIS is a project for letting AI assistants and test suites operate Android phones and emulators from natural-language instructions. It supports cross-app workflows, such as navigating Google Maps and YouTube, and locates targets through element indices, coordinates, OCR, accessibility hierarchies, and visual fallbacks for custom interfaces. The system provides a reactive Flash profile that typically takes about 3–5 seconds per step and a Pro profile that adds planning, pre-execution checks, recovery handling, checkpoint verification, diagnostics, and optional reports, with typical steps taking about 15–40 seconds. ARTEMIS reports a 99%+ task-completion rate on Google Research’s AndroidWorld benchmark, which covers more than 20 apps and over 100 multi-step tasks. Through its native Model Context Protocol server, AI development environments including Antigravity, Claude Code, Codex, and Windsurf can control connected devices, inspect state, collect Logcat output and screenshots, and produce structured testing results. The project also offers a web console with live screen projection and execution replay, a command-line interface, and a zero-runtime-dependency Python SDK designed for integration with frameworks such as pytest and CI/CD pipelines. A connected Android device must have USB debugging enabled, or an emulator must be available; the startup scripts can install ADB, scrcpy, FFmpeg, and Python dependencies. The first device task installs an Artemis Accessibility Helper that reads screen layout locally and sends nothing elsewhere, with UIAutomator2 available as a fallback or alternative. Flash lacks Pro’s planning and verification features, while Pro is intended for longer workflows and exploratory testing. The repository is licensed under Apache 2.0 and lists Android Studio integration, iOS support, on-device lightweight vision models, and voice interaction as roadmap items.