DeepMind’s new AI model runs natively on robots for faster performance
Google DeepMind is releasing an on-device version of its Gemini Robotics AI model that functions without requiring an internet connection. This vision-language-action model (VLA) includes the same dexterous capabilities introduced in March, but Google notes it is now compact and efficient enough to run directly on a robot.
The flagship Gemini Robotics model enables robots to perform diverse physical tasks, even those they weren’t specifically trained for. It supports generalizing new scenarios, understanding and reacting to instructions, and carrying out activities that involve fine motor control.
According to Carolina Parada, head of robotics at Google DeepMind, the original Gemini Robotics model uses a hybrid method that works both on-device and in the cloud. With this new device-only model, however, users can access offline capabilities that offer nearly the same performance as the flagship version.

Apptronik’s Apollo humanoid bot followed by Google’s ALOHA system. GIF: GoogleThe on-device model can handle multiple tasks immediately and adapt to unfamiliar situations “with as few as 50 to 100 demonstrations,” Parada explains. Although Google initially trained the model only on its ALOHA robot, the company successfully adapted it to other robotic platforms, such as Apptronik’s humanoid Apollo robot and the dual-arm Franka FR3 robot.
“The Gemini Robotics hybrid model remains more powerful, but we've been genuinely impressed by the performance of this on-device version,” Parada adds. “It can be seen as an entry-level model or an ideal solution for environments with unreliable internet.” It also suits organizations with strict security protocols.
In addition to the new model, Google is making available a software development kit (SDK) for developers to test and customize it — the first time such a toolkit has been released for one of Google DeepMind’s VLAs.
The on-device Gemini Robotics model and its accompanying SDK will initially be provided to a select group of trusted testers while Google continues addressing and minimizing potential safety concerns.
Related article
Google Tests Remy AI Agent for Gemini as Focus Shifts to User Control
According to Business Insider, Google is testing Remy, a new AI personal agent for Gemini. This tool aims to execute tasks on behalf of users, streamlining both professional workflows and daily routines.Currently, Remy is undergoing testing in an int
Ollie bets privacy focus to win AI assistant race
To be genuinely helpful, an AI assistant must understand its user deeply. Ollie, a personal assistant designed for daily life, operates on the premise that this doesn’t require surrendering your data or compromising your privacy.While certain enterpr
How AI LIVE: London Will Explore AI & Industrial Automation
The summit will convene C-suite executives from around the globe to address pressing challenges in global industries, ranging from AI-driven disruption to economic volatility.AI LIVE: The London Summit will gather over 2,000 international leaders und
Related Special Topic Recommendations
Comments (1)
0/500
That's a huge step forward for robotics! On-device processing is finally catching up, but I'm kinda wondering how expensive the hardware will be - can small startups afford it? 🤔 Also, curious if offline functionality means it's less likely to be updated or influenced remotely, which might be a double-edged sword.
Google DeepMind is releasing an on-device version of its Gemini Robotics AI model that functions without requiring an internet connection. This vision-language-action model (VLA) includes the same dexterous capabilities introduced in March, but Google notes it is now compact and efficient enough to run directly on a robot.
The flagship Gemini Robotics model enables robots to perform diverse physical tasks, even those they weren’t specifically trained for. It supports generalizing new scenarios, understanding and reacting to instructions, and carrying out activities that involve fine motor control.
According to Carolina Parada, head of robotics at Google DeepMind, the original Gemini Robotics model uses a hybrid method that works both on-device and in the cloud. With this new device-only model, however, users can access offline capabilities that offer nearly the same performance as the flagship version.

The on-device model can handle multiple tasks immediately and adapt to unfamiliar situations “with as few as 50 to 100 demonstrations,” Parada explains. Although Google initially trained the model only on its ALOHA robot, the company successfully adapted it to other robotic platforms, such as Apptronik’s humanoid Apollo robot and the dual-arm Franka FR3 robot.
“The Gemini Robotics hybrid model remains more powerful, but we've been genuinely impressed by the performance of this on-device version,” Parada adds. “It can be seen as an entry-level model or an ideal solution for environments with unreliable internet.” It also suits organizations with strict security protocols.
In addition to the new model, Google is making available a software development kit (SDK) for developers to test and customize it — the first time such a toolkit has been released for one of Google DeepMind’s VLAs.
The on-device Gemini Robotics model and its accompanying SDK will initially be provided to a select group of trusted testers while Google continues addressing and minimizing potential safety concerns.
Ollie bets privacy focus to win AI assistant race
To be genuinely helpful, an AI assistant must understand its user deeply. Ollie, a personal assistant designed for daily life, operates on the premise that this doesn’t require surrendering your data or compromising your privacy.While certain enterpr
How AI LIVE: London Will Explore AI & Industrial Automation
The summit will convene C-suite executives from around the globe to address pressing challenges in global industries, ranging from AI-driven disruption to economic volatility.AI LIVE: The London Summit will gather over 2,000 international leaders und
That's a huge step forward for robotics! On-device processing is finally catching up, but I'm kinda wondering how expensive the hardware will be - can small startups afford it? 🤔 Also, curious if offline functionality means it's less likely to be updated or influenced remotely, which might be a double-edged sword.





Home






