Google DeepMind Introduces AI Model That Runs Locally on Robots | PYMNTS.com | Goose Pod

Authors: PYMNTS

Publisher:

PYMNTS.com

Published: 6/24/2025

Language:English

--:--

David

Good morning Wang Kang, I'm David, and with me is the always insightful Emiky. Welcome to Goose Pod for you! Today is Sunday, June 29th, and we're diving into a fascinating topic: Google DeepMind's new AI model that runs locally on robots. This is a game-changer, and I'm eager to get started.

Emiky

Good morning, David, and to all our listeners! I'm so excited about this one. Imagine robots that don't need constant Wi-Fi to be super smart. It’s like giving them their own portable brain! We're talking about real autonomy here, which is just incredible for the future of robotics.

David

Precisely, Emiky. Google DeepMind has introduced what they call the Gemini Robotics On-Device model. The striking aspect, as you highlighted, is its ability to operate entirely locally on robotic devices. This means no reliance on a data network, which is a significant leap for latency-sensitive applications and ensures robustness even in environments with intermittent or zero connectivity. It's a foundational shift in how robots interact with their environment.

Emiky

Right, so think about it like this: your phone can do amazing things even when you're offline, because the apps run right on the device. Now, robots can have that same kind of independence! This model, Gemini Robotics On-Device, gives them general-purpose dexterity and allows for super-fast task adaptation. It's designed for bi-arm robots, which are those cool robots with two arms, and it enables really quick experimentation with complex manipulations.

David

And the tasks it can perform are quite impressive, demonstrating its versatility. We're talking about robots unzipping bags, folding clothes, zipping a lunchbox, drawing a card, pouring salad dressing, and even assembling products. These aren't simple, repetitive factory tasks; they require a high degree of nuanced understanding and fine motor control, all guided by natural language instructions. It's truly a testament to the model's advanced capabilities.

Emiky

It really is! And to understand how we got here, we need to look at the 'background'. This new on-device model actually builds on Google DeepMind's previous work with Gemini Robotics, which they introduced back in March. Think of Gemini Robotics as the brilliant older sibling, and Gemini Robotics On-Device is like the younger sibling who's even more optimized and independent. It leverages the multimodal capabilities of the broader Gemini platform, which processes text, images, and audio.

David

Indeed, Emiky. The strategic decision to make Gemini multimodal has been pivotal, paving the way for enhanced reasoning capabilities in AI, which directly translates to more intelligent robotic control. The impetus for this on-device model stems from a clear need for robotic systems that can operate reliably in diverse environments, especially where network access is limited or nonexistent. This includes remote locations, disaster zones, or even within manufacturing facilities where low latency is critical for safety and efficiency.

Emiky

Exactly, it's about making robots more robust and responsive! The development timeline shows a clear progression. Back in December 2024, Apptronik, a company that makes humanoid robots, partnered with Google DeepMind. Then in March 2025, they launched the original Gemini Robotics models. And now, just this June, we have the optimized On-Device version. It's a fast-paced evolution, showing how quickly they're pushing the boundaries.

David

This partnership with Apptronik, specifically involving their Apollo humanoid robot, is a crucial real-world application. It demonstrates the model's adaptability across various hardware platforms. The core motivation, as Carolina Parada, Google DeepMind’s Senior Director of Robotics, articulated, is to ensure robustness in environments with intermittent or zero connectivity, which is a common challenge for many industrial and logistical applications. It's about bringing advanced AI directly to the point of action.

Emiky

And it makes so much sense! Imagine a robot in a remote warehouse, or even helping out in your home, without needing a super strong Wi-Fi signal all the time. That's the dream, right? This VLA, or Vision Language Action model, basically allows the robot to see, understand your language, and then act on it directly. It’s like giving the robot its own internal interpreter that never gets disconnected. It’s a huge step for practical, everyday robotic use.

David

While the capabilities are impressive, it's important to discuss the 'conflict' or challenges that this technology aims to overcome. One significant hurdle in robotics has always been latency. When a robot relies on cloud processing, even a millisecond of delay can be critical, especially for precise or safety-sensitive tasks. Operating locally eliminates this bottleneck, providing near-instantaneous responses, which is a major technical triumph.

Emiky

Totally! Think of it like trying to play a video game with terrible internet lag – it's impossible to be precise. Robots doing delicate tasks, like assembling a tiny part or pouring a drink, need that instant feedback. Another 'conflict' is the sheer complexity of making AI models general-purpose. Previously, robots were often programmed for one specific task. But with natural language understanding, the goal is to make them adaptable, which is super hard!

David

And this is where the fine-tuning capability of Gemini Robotics On-Device truly shines. Many AI models require vast datasets for adaptation. However, this model can quickly adapt to new tasks with as few as 50 to 100 demonstrations. This addresses a common challenge in AI deployment, as it significantly reduces the data and computational resources needed for customization. This 'low-shot learning' is a critical advancement for practical implementation.

Emiky

It's like a robot that's a quick learner! You show it a few times, and it gets it, instead of needing a million examples. Another 'conflict' in the robotics world is how crowded the market is getting. Lots of companies are trying to develop AI-powered robots that can do general tasks. So, Google DeepMind isn't just innovating; they're also navigating a very competitive landscape. Their on-device approach gives them a unique edge.

David

Absolutely. The market is indeed becoming competitive, with many players pushing the boundaries of what robots can achieve. This makes innovation like local processing and rapid adaptation even more crucial for differentiation. Furthermore, the inherent variability of real-world environments presents another layer of conflict, requiring robust models that can generalize effectively from controlled demonstrations to unpredictable scenarios, which the On-Device model aims to address.

Emiky

So, what's the 'impact' of all this? Well, the most immediate impact is that robots become way more reliable and usable in places where internet connectivity is a problem. Think about factories, warehouses, or even remote areas. This means businesses can deploy advanced robotics without worrying about network stability, which can dramatically lower operational costs.

David

Precisely. Experts are viewing this as a significant step towards making powerful robotics models more accessible and adaptable. The on-device functionality and low-shot learning characteristics are expected to reduce deployment barriers for businesses, fostering the widespread adoption of robotic technology across various sectors. This includes manufacturing, logistics, and security, ultimately leading to more robust and responsive robotic systems. It really redefines the landscape of robot applications.

Emiky

It's like robots going from needing a constant tether to being truly free-roaming and smart on their own. This could open up a whole new genre of consumer products for Google, too! Imagine an intelligent home assistant robot that can actually fold your laundry, or help with dinner prep, without needing to 'call home' to the cloud every five seconds. The privacy implications are also positive, as less data needs to be transmitted.

David

Looking ahead to the 'future,' this development points towards a continued push for increased autonomy and adaptability in robotic systems. Google DeepMind is also releasing a Gemini Robotics SDK, a software development kit. This will allow developers to evaluate the On-Device model and quickly adapt it to new domains, accelerating innovation by enabling customization for specific applications. It democratizes access to advanced robotics.

Emiky

That's awesome! It means more people can experiment and build cool stuff with these robots. It’s like giving everyone the tools to create their own smart robot helpers. And DeepMind is also really focusing on responsible AI development, which is super important. They're making sure these powerful models are deployed ethically and safely, which is a great sign for the future.

David

Indeed. Their emphasis on a holistic safety approach, including semantic and content safety, alongside physical safety measures, is paramount. This ensures that as robots become more capable and autonomous, their deployment is carefully considered for societal impact and minimized risks. The ability to fine-tune with limited demonstrations also suggests a future where robots can rapidly learn and adapt to entirely new tasks and environments, leading to wider adoption in diverse industries.

Emiky

Wow, what a journey into the future of robotics! From offline brains to quick learners, Gemini Robotics On-Device is definitely shaking things up. Thanks for joining us on this exciting discussion, David!

David

It was a pleasure, Emiky. That's the end of today's discussion. Thank you for listening to Goose Pod, Wang Kang. See you tomorrow, for more fascinating insights!

# News Summary: Google DeepMind Introduces AI Model That Runs Locally on Robots * **News Title**: Google DeepMind Introduces AI Model That Runs Locally on Robots * **Report Provider/Author**: PYMNTS.com / PYMNTS * **Date/Time Period Covered**: Published on June 24, 2025. The article references related developments and reports from March, April, and February of the same year. * **News Type**: Technology, Artificial Intelligence (AI), Robotics, Digital Transformation, Innovation. --- ## Main Findings and Conclusions Google DeepMind has unveiled a new **vision language action (VLA) model** named **Gemini Robotics On-Device**, designed to operate directly on robotic devices without requiring an internet connection. This advancement signifies a step towards more robust and responsive robotic systems, particularly for applications where network connectivity is unreliable or latency is critical. ## Key Features and Capabilities * **Local Operation**: The model runs entirely on the robotic device, eliminating the need for data network access. This is crucial for "latency sensitive applications and ensures robustness in environments with intermittent or zero connectivity," according to Carolina Parada, Google DeepMind's Senior Director and Head of Robotics. * **General-Purpose Dexterity**: Gemini Robotics On-Device is engineered for broad manipulation capabilities and rapid adaptation to new tasks. * **Bi-Arm Robot Focus**: The model is specifically designed for use with bi-arm robots, facilitating advanced dexterous manipulation. * **Natural Language Understanding**: It can follow instructions given in natural language, enabling intuitive control. * **Task Versatility**: The model demonstrates proficiency in a range of complex tasks, including: * Unzipping bags * Folding clothes * Zipping a lunchbox * Drawing a card * Pouring salad dressing * Assembling products * **Fine-Tuning Capability**: This is Google DeepMind's first VLA model that is available for fine-tuning by developers. This allows for customization and improved performance for specific applications. * **Rapid Task Adaptation**: The model can quickly adapt to new tasks with "as few as 50 to 100 demonstrations," showcasing its strong generalization capabilities from foundational knowledge. ## Context and Market Trends * **Building on Previous Work**: Gemini Robotics On-Device builds upon the capabilities of Gemini Robotics, which was initially introduced in March. * **Industry Shift**: The development aligns with a broader trend in Silicon Valley where large language models are being integrated into robots, enabling them to comprehend natural language commands and execute complex tasks. * **Multimodality of Gemini**: Google's strategic decision to make Gemini multimodal (processing and generating text, images, and audio) is highlighted as a path toward enhanced reasoning capabilities, potentially leading to new consumer products. * **Crowded Market**: The field of AI-powered robots capable of general tasks is becoming increasingly competitive, with several other companies also making significant advancements. ## Key Personnel Quotes * **Carolina Parada (Senior Director and Head of Robotics, Google DeepMind)**: * "Since the model operates independent of a data network, it’s helpful for latency sensitive applications and ensures robustness in environments with intermittent or zero connectivity." * "While many tasks will work out of the box, developers can also choose to adapt the model to achieve better performance for their applications." * "Our model quickly adapts to new tasks, with as few as 50 to 100 demonstrations — indicating how well this on-device model can generalize its foundational knowledge to new tasks."

Google DeepMind Introduces AI Model That Runs Locally on Robots | PYMNTS.com

Read original at PYMNTS.com →

Google DeepMind introduced a vision language action (VLA) model that runs locally on robotic devices, without accessing a data network.The new Gemini Robotics On-Device robotics foundation model features general-purpose dexterity and fast task adaptation, the company said in a Tuesday (June 24) blog post.

“Since the model operates independent of a data network, it’s helpful for latency sensitive applications and ensures robustness in environments with intermittent or zero connectivity,” Google DeepMind Senior Director and Head of Robotics Carolina Parada said in the post.Building on the task generalization and dexterity capabilities of Gemini Robotics, which was introduced in March, Gemini Robotics On-Device is meant for bi-arm robots and is designed to enable rapid experimentation with dexterous manipulation and adaptability to new tasks through fine-tuning, according to the post.

The model follows natural language instructions and is dexterous enough to perform tasks like unzipping bags, folding clothes, zipping a lunchbox, drawing a card, pouring salad dressing and assembling products, per the post.It is also Google DeepMind’s first VLA model that is available for fine-tuning, per the post.

“While many tasks will work out of the box, developers can also choose to adapt the model to achieve better performance for their applications,” Parada said in the post. “Our model quickly adapts to new tasks, with as few as 50 to 100 demonstrations — indicating how well this on-device model can generalize its foundational knowledge to new tasks.

”Google DeepMind’s Gemini Robotics is one of several companies’ efforts to develop humanoid robots that can do general tasks, PYMNTS reported in March.Robotics are in fashion as in Silicon Valley as large language models are giving robots the capability to understand natural language commands and do complex tasks.

The company’s advancements in Gemini Robotics show that the decision to make Gemini multimodal — taking and generating text, images and audio — is the path toward better reasoning. Gemini’s multimodality can spawn a whole new genre of consumer products for Google, PYMNTS reported in April.Several other companies are also developing AI-powered robots demonstrating advancements in general tasks, making for a crowded market, PYMNTS reported in February.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.See More In: artificial intelligence, deepmind, digital transformation, GenAI, Google, Innovation, News, PYMNTS News, Robots, Technology, What's Hot

Analysis

Phenomenon+

Conflict+

Background+

Impact+

Future+

Related Podcasts

Amazon founder Jeff Bezos says ‘millions of people’ will be living in space by 2045—and robots will commute on our behalf to the moon | Fortune

12/4/2025

A $5.4 Billion Robotics Sale Signals Europe’s Loss of Control

11/12/2025

A delivery robot collided with a disabled man on L.A. street. The aftermath is getting ugly

10/22/2025

View All Podcasts →