Overview

• The model runs entirely on-device with minimal compute resources for bi-arm robots, ensuring low-latency performance in environments with intermittent or no connectivity.
• It executes out-of-the-box dexterous actions—such as unzipping bags and folding clothes—demonstrating strong general-purpose manipulation skills.
• In benchmarks, the on-device model closely matches the cloud-based hybrid Gemini Robotics version and outperforms other on-device AI on challenging out-of-distribution and multi-step tasks.
• Access is limited to a trusted tester program, and Google is providing a Gemini Robotics SDK for developers to evaluate, fine-tune, and deploy the model in both simulation and real-world settings.
• Because the on-device variant lacks built-in semantic safety reasoning, developers are advised to implement their own safety systems following Google’s layered safety framework.