.svg){kind=spacer}
Details
- Google DeepMind has introduced Gemini Robotics On-Device, its first vision–language–action model that operates entirely on robot hardware without needing an internet connection.
- This model maintains the advanced dexterity of the cloud-based Gemini Robotics but is streamlined to suit the computing and memory limits of common humanoid and dual-arm industrial robots.
- Out-of-the-box capabilities enable complex two-handed tasks, and the system can quickly learn new duties with just 50–100 human demonstrations.
- Though trained on the ALOHA dataset, the platform adapts to various robot bodies by following instructions given in natural language.
- The new Gemini Robotics SDK lets developers fine-tune, test, and validate robot control policies using the MuJoCo physics simulator prior to real-world use.
- Processing directly on the robot reduces latency and bandwidth demands, makes robots less reliant on the cloud, and supports applications in locations with strict data requirements such as warehouses and remote sites.
Impact
This move challenges competitors like Nvidia’s Isaac Sim and Tesla’s Optimus, which still depend on cloud services for complex operations. By eliminating ongoing cloud costs, it could boost advanced robotics adoption among businesses previously limited by cost or regulatory compliance. On-device processing also better aligns with data-localization rules, positioning DeepMind as a leader in privacy-friendly, scalable robotics solutions.
