Offline AI boosts robots’ capabilities without the cloud
Google DeepMind has unveiled Gemini Robotics On-Device, a new AI model that enables robots to perform complex tasks locally without relying on an internet connection. The announcement marks a significant leap toward autonomous robotic systems capable of operating in real time using onboard computation.
Building on the cloud-based Gemini Robotics model introduced in March, this version allows developers to control robot behavior through natural language prompts and fine-tune responses to fit specific applications. Google says the new model matches or nearly matches the performance of its cloud-based predecessor and outperforms competing on-device models, although no direct comparisons were disclosed.
Demonstrations show practical, real-world capabilities
In a demonstration, Google showcased robots using the new model to unzip bags, fold clothes, and perform precise assembly tasks on industrial belts. Notably, a bi-arm Franka FR3 robot successfully executed tasks it had never encountered before, including handling unfamiliar objects and environments.
Though initially trained for ALOHA robots, the model adapted effectively to the Franka FR3 and Apptronik’s Apollo humanoid robot, underscoring its versatility across hardware platforms.
New SDK empowers developers with fast training tools
Alongside the model, Google is launching the Gemini Robotics SDK, giving developers the tools to train robots using only 50 to 100 demonstrations per task within the MuJoCo physics simulator. This simplified training process could drastically reduce development time and data requirements for robotics applications.
Competition in robotics AI heats up
Google’s move comes as several major players intensify efforts to integrate AI into robotics. Nvidia is developing a full-stack platform for humanoid robot foundation models, while Hugging Face is advancing open-source robotics datasets and models. Meanwhile, South Korea’s RLWRLD, backed by Mirae Asset, is also building foundational models for autonomous machines.
With the launch of Gemini Robotics On-Device, Google DeepMind is staking a claim in the emerging frontier of offline, general-purpose robotic intelligence — a field poised to reshape industries from manufacturing to domestic automation.
