News

Google Unveils On-Device AI Model for Robots

• By John K. Waters
• 06/24/2025

Google has released an artificial intelligence model designed to run directly on robotic devices without requiring internet connectivity, the company announced Tuesday.

The new software, called Gemini Robotics On-Device, represents Google's latest effort to bring AI capabilities to physical robots while addressing connectivity and latency challenges that have limited robotic applications in industrial and commercial settings.

The model builds on Google's earlier Gemini Robotics platform, introduced in March, but has been optimized to operate locally on robot hardware rather than relying on cloud-based processing. This approach could prove crucial for robotics applications in environments with poor internet connectivity or where split-second response times are essential.

Technical Capabilities
According to Google, the new model can perform complex manipulation tasks including unzipping bags and folding clothes while following natural language instructions. The company demonstrated the technology on bi-arm robots performing what it describes as "dexterous manipulation" tasks.

In testing, Google found the model could adapt to new tasks using as few as 50 to 100 demonstration examples, suggesting the technology could be quickly customized for specific industrial or commercial applications.

The company also showed the model working across different robot designs, including ALOHA research robots, Franka FR3 industrial arms, and Apptronik's Apollo humanoid robot. This cross-platform compatibility could reduce barriers to adoption across the robotics industry.

Industry Context
The release comes as companies across multiple sectors are exploring robotics applications amid labor shortages and increasing automation needs. However, many current robotic systems require constant internet connectivity to access AI processing power housed in remote data centers.

Google's on-device approach addresses two key limitations of cloud-dependent robotics: network latency that can slow robot responses, and connectivity requirements that limit where robots can operate effectively.

The company plans to make the model available initially to a select group of developers through what it calls a "trusted tester program." Participants will receive access to both the AI model and a software development kit designed to help adapt the technology to specific applications.

Safety Measures
Google said it has implemented multiple safety protocols for the robotics model, including content filtering and integration with low-level safety controllers that govern robot actions. The company's internal safety teams are conducting ongoing assessments of the technology's real-world implications.

The announcement reflects broader industry efforts to move AI processing closer to where it's needed, reducing dependence on centralized computing resources while improving response times for time-sensitive applications.

Companies including Tesla, Boston Dynamics, and numerous startups are developing AI-powered robots for applications ranging from warehouse automation to household assistance, though widespread commercial deployment has been limited by technical and cost constraints.

Google's parent company Alphabet has been investing heavily in robotics research for over a decade, though previous efforts including acquisitions of robotics companies have yielded limited commercial products.