Google DeepMind launches Gemini 'Robotics On-Device'

Google DeepMind has today released Gemini Robotics On-Device, a localised version of its robotics AI designed to run directly on physical hardware.

The launch builds on the Gemini Robotics platform that was introduced earlier this year, which added real-world manipulation and reasoning to Gemini 2.0’s multimodal AI capabilities.

The 'on-device' model brings general-purpose dexterity and task generalisation to robots while operating independently of cloud services, Google explained in a statement.

DeepMind said the model, running on the robot hardware locally, is suitable for applications where connectivity is unreliable or latency needs to be minimised.

Engineered for use with bi-arm robots, the new system can follow natural language commands and perform multi-step tasks such as folding clothes or unzipping bags.

It can also be fine-tuned with as few as 50 demonstrations and adapted across various robot types, it added.

Developers interested in testing the model can gain access via DeepMind’s trusted tester programme, Google highlighted. A new software development kit and simulated environments are also available to accelerate research and application in the field.