In 2025-03, Google DeepMind introduced a pair of artificial intelligence models aimed at the physical world of robotics. The two systems, named Gemini Robotics and Gemini Robotics-ER, were presented as a significant step forward in giving machines a more nuanced touch and a better grasp of their surroundings. The announcement, made on a Wednesday, positioned these models as a kind of cognitive layer for robots of various sizes and configurations, from industrial arms to humanoid platforms.
The most striking demonstrations involved tasks that require a light hand. According to the information released by DeepMind, the Gemini Robotics model can fold delicate origami and close zipper bags without causing damage. These are not trivial operations for a machine. Origami, in particular, demands precise pressure control, an understanding of paper's flexibility, and the ability to adjust in real time if a fold goes slightly off. Zipper bags, while seemingly simpler, require a steady grip and careful alignment to avoid tearing the plastic or jamming the closure. That the model can handle both suggests a level of fine motor skill that has historically been difficult to achieve in robotics.
The underlying capability here is what DeepMind calls generalization. In simple terms, this is the ability of an AI system to perform a task it was not explicitly trained for. Many existing robot control systems are brittle: they excel at the specific scenarios they were programmed for, but falter when faced with a new object, a different layout, or an unexpected variable. Gemini Robotics, by contrast, is claimed to demonstrate much stronger generalization than its predecessors. In the company's own testing, the model reportedly "more than doubles performance on a comprehensive generalization benchmark compared to other state-of-the-art vision-language-action models." That is a notable claim, though the specific benchmark details and the full list of competing models were not fully disclosed in the source material.
Vision-language-action models, or VLAs, are a current frontier in robot learning. They combine visual input (what the robot sees), language understanding (what the robot is instructed to do), and action output (how the robot moves). By integrating these three streams, a VLA can theoretically interpret a command like "pick up the red cup" and execute it, even if it has never seen that particular cup before. Gemini Robotics is built on this architecture, but with an added emphasis on physical interaction. The model is designed not just to recognize objects, but to understand how to handle them—how much force to apply, how to orient a gripper, how to respond if an object shifts.
The second model, Gemini Robotics-ER, appears to be focused on embodied reasoning. While the source material does not provide exhaustive technical specifications, the "ER" designation suggests a model that can reason about spatial relationships, physical constraints, and the consequences of actions in a three-dimensional environment. This would be the component that allows a robot to plan a path around an obstacle, or to understand that a fragile object requires a gentler approach than a sturdy one.
One concrete application of this technology is already in motion. Google DeepMind announced that Gemini Robotics will serve as the "robot brain" for Apptronik's Apollo humanoid robot. Apollo is a bipedal robot designed for real-world work, and pairing it with a more capable AI system could accelerate its usefulness in settings that require both mobility and dexterity. The source material does not specify a timeline for this integration, nor does it detail the commercial terms of the arrangement. What is clear is that Google is not just publishing research papers; it is actively seeking to place its AI inside commercially available hardware.
It is worth noting what the source material does not say. There are no figures for the number of robots that will use Gemini Robotics, no pricing information, no availability dates for developers or enterprises, and no details on the computational requirements to run the model. The announcement is heavy on capability claims and light on deployment specifics. For a robotics industry that has grown accustomed to bold AI announcements followed by long integration timelines, this is a familiar pattern. The technology is real, the demonstrations are compelling, but the path from lab to warehouse floor is often longer than the press release suggests.
Why it matters for European robot service
For the European robotics ecosystem, the arrival of Gemini Robotics is significant for several reasons, even if the model itself is developed on the other side of the Atlantic.
First, consider the state of the European robot service market. The region has a strong tradition of industrial automation, particularly in automotive manufacturing, logistics, and precision engineering. European companies have been early adopters of collaborative robots, or cobots, which are designed to work alongside humans rather than replace them. These cobots are often deployed in small and medium-sized enterprises (SMEs) that need flexibility more than raw speed. The ability to reprogram a robot for a new task without extensive retraining is a major value proposition for these businesses. Gemini Robotics, with its emphasis on generalization, speaks directly to that need. If a robot can learn to fold paper and close zippers, it can likely be adapted to handle a wider variety of assembly, packaging, and inspection tasks than current systems allow.
Second, the humanoid angle is particularly relevant for Europe. Several European startups and research institutes are working on humanoid robots, and the idea of a shared "brain" that can be plugged into different bodies is an attractive one. If Google's model can indeed serve as a versatile cognitive layer, it could lower the barrier to entry for smaller European robot manufacturers who lack the resources to develop their own advanced AI. Instead of building a proprietary control system from scratch, they could integrate a model like Gemini Robotics and focus their engineering efforts on hardware design, safety systems, and application-specific tooling. This could accelerate the commercialization of European humanoid robots, which have so far struggled to move beyond the prototype stage.
Third, there is a broader question of strategic autonomy. Europe has been increasingly vocal about reducing its dependence on non-European technology in critical sectors, and robotics is no exception. The European Commission has funded numerous projects aimed at developing homegrown AI and robotics capabilities. The arrival of a powerful US-based model like Gemini Robotics creates both an opportunity and a challenge. On one hand, it provides European companies with access to world-class AI without requiring them to match Google's research budget. On the other hand, it raises concerns about dependency. If European robots run on American AI, what happens if the technology is restricted, or if the terms of use change? The source material does not address these geopolitical dimensions, but they are impossible to ignore for anyone tracking the industry.
Fourth, the fine motor skills demonstrated by Gemini Robotics have direct implications for European service sectors. Consider healthcare and eldercare, where robots are being explored as assistants for tasks like preparing meals, helping with dressing, or handling medical instruments. These tasks require exactly the kind of delicate touch that the model appears to offer. Similarly, in the food industry, robots that can handle soft or fragile items—pastries, fruits, prepared dishes—without crushing them have long been a goal. The origami and zipper demonstrations suggest that the model has the dexterity to handle such items, though the source material does not provide specific examples in these domains.
Fifth, the generalization benchmark claim deserves attention from a European perspective. European robotics researchers have been active in developing benchmarks for robot learning, and the claim that Gemini Robotics more than doubles the performance of other state-of-the-art VLAs is a strong statement. If independently verified, it would represent a significant leap forward. However, the source material does not provide details on the benchmark methodology, the tasks included, or the specific comparison models. European buyers and researchers should approach such claims with a healthy dose of skepticism until more information is available. Benchmark results can be influenced by task selection, evaluation protocols, and even the specific hardware used for testing.
Finally, the Apptronik partnership signals that Google is serious about commercial deployment, not just research. For European companies, this means that advanced AI for robots may soon be available through commercial channels, potentially as a service or a licensed model. This could be a game-changer for European robot integrators who currently rely on less capable, more rigid control systems. The ability to upgrade a robot's "brain" without replacing its body is an attractive proposition for end users who have already invested in hardware.
What buyers and operators should know
For buyers and operators of robot services in Europe, the announcement of Gemini Robotics raises several practical considerations. It is important to separate the demonstrated capabilities from the aspirational claims, and to understand what the technology can and cannot do based on the available information.
First, the demonstrated tasks—folding origami and closing zipper bags—are impressive but narrow. They show that the model can handle delicate materials and perform precise manipulations. However, they do not prove that the model can handle the full range of tasks that a commercial robot might face in a factory, warehouse, or hospital. The source material does not provide examples of the model performing industrial tasks such as assembly, welding, or material handling. Buyers should therefore view the demonstrations as proof of concept for fine motor skills, not as evidence that the model is ready for all applications.
Second, the generalization claim is significant but needs scrutiny. The source material states that Gemini Robotics "more than doubles performance on a comprehensive generalization benchmark compared to other state-of-the-art vision-language-action models." This is a quantitative claim, but the specifics are not disclosed. Which benchmark was used? What tasks did it include? How many models were compared? Without these details, it is difficult to assess the validity of the claim. Buyers should ask for more information from Google or DeepMind before making procurement decisions based on this metric.
Third, the integration with Apptronik's Apollo humanoid is a notable development, but it is not yet a commercial product. The source material does not state when Apollo robots with Gemini Robotics will be available, how much they will cost, or which markets will be served first. European buyers interested in humanoid robots should monitor this partnership closely but should not expect immediate availability. The timeline from announcement to deployment in the robotics industry is often measured in years, not months.
Fourth, there are unanswered questions about the hardware requirements. Running a large vision-language-action model requires significant computational resources. The source material does not specify whether Gemini Robotics can run on edge devices, such as the onboard computers typically found in robots, or whether it requires cloud connectivity. This is a critical consideration for European operators, particularly those in environments with limited internet connectivity or strict data privacy requirements. If the model requires cloud processing, that raises questions about latency, reliability, and data security. The source material does not address these issues.
Fifth, the source material does not mention safety certifications or compliance with European regulations. Robots deployed in the EU must meet strict safety standards, including the Machinery Directive and, increasingly, AI-specific regulations under the proposed EU AI Act. The source material provides no information on whether Gemini Robotics has been tested for compliance with these standards. Buyers should not assume that a model developed by a US company automatically meets European regulatory requirements. This is an area where more information is needed before any procurement decisions are made.
Sixth, the source material does not disclose pricing or licensing terms. It is unclear whether Gemini Robotics will be offered as a cloud API, a downloadable model, or a pre-integrated solution with specific robot manufacturers. Each of these models has different implications for cost, control, and customization. European buyers should seek clarity on these terms before committing to any partnership or purchase.
Seventh, there is the question of ongoing support and updates. AI models are not static; they require regular updates, retraining, and maintenance. The source material does not describe Google's plans for supporting Gemini Robotics over time, nor does it indicate how improvements will be rolled out to existing users. For European operators who rely on robots for critical operations, the long-term viability of the AI model is as important as its initial capabilities.
Eighth, the source material does not address the potential for bias or errors in the model. Like all AI systems, Gemini Robotics is likely to have limitations and failure modes. The source material does not describe any testing for edge cases, adversarial scenarios, or unexpected environments. Buyers should be aware that no AI system is perfect, and they should plan for contingencies in case the model fails to perform as expected.
Ninth, the source material does not mention any partnerships or integrations with European robot manufacturers beyond the Apptronik deal. It is unclear whether Google plans to work with European companies directly, or whether European buyers will need to go through intermediaries. This could affect availability, support, and pricing in the European market.
Tenth, and finally, the source material does not provide any information on the environmental impact of running Gemini Robotics. Large AI models consume significant energy, both during training and during inference. For European companies with sustainability targets, this is a relevant consideration. The source material is silent on this topic, so buyers will need to seek information from other sources.
In summary, Gemini Robotics is a promising development with demonstrated capabilities in fine motor skills and generalization. However, the source material leaves many practical questions unanswered. European buyers and operators should approach the technology with informed optimism, seeking additional details on benchmarks, hardware requirements, regulatory compliance, pricing, and support before making any commitments. The technology has the potential to transform robot services in Europe, but the path from announcement to deployment is still unclear.
Sources
https://arstechnica.com/ai/2025/03/googles-origami-folding-ai-brain-may-power-new-wave-of-humanoid-robots/
Published by Vigla Media OÜ (Estonia).