Google Launches European Robotics Accelerator: 15 Startups Selected as the Physical AI Race Begins

Google Bets on Physical AI with the Launch of Its European Robotics Accelerator

Google recently announced the official launch of its Robotics Accelerator program, with the first cohort of 15 European startups selected. This three-month accelerator will provide participating companies with access to Google's AI technology stack, Gemini Robotics models, and hands-on support from Google's internal teams, all aimed at advancing "Physical AI" development in Europe.

This move marks a deepening of Google's strategic positioning in the Embodied AI space and reflects the intense interest global tech giants are showing in the convergence of robotics and AI.

What Is Physical AI? Why Is It the Industry's New Focus?

Physical AI is a concept that has garnered significant attention in the AI field in recent years. It refers to artificial intelligence systems capable of perceiving, understanding, and taking actions in the real physical world. Unlike purely digital AI (such as chatbots or image generation tools), Physical AI must simultaneously tackle a far more complex set of engineering challenges, including robot control, environmental perception, and real-time decision-making.

The core technical challenge of Physical AI lies in bridging the "sim-to-real gap" — the chasm between simulation and reality. AI models trained in digital environments face a host of unmodeled factors when deployed in the real physical world: sensor noise, object deformation, lighting variations, and dynamic uncertainties. Traditional robots rely on precise mathematical modeling and pre-programmed control strategies, whereas Physical AI attempts to endow robots with generalization capabilities through end-to-end learning, reinforcement learning, and large-scale data-driven approaches — enabling them to act reasonably even in previously unseen scenarios. Behind this paradigm shift is the successful transfer of the Transformer architecture to the robotics domain, along with the continued accumulation of large-scale robot manipulation datasets (such as Google's RT-2 and Open X-Embodiment).

From NVIDIA CEO Jensen Huang repeatedly emphasizing that "Physical AI is the next wave" to Google establishing a dedicated accelerator program, an industry consensus is forming: The capabilities of large language models are spilling over from the digital world into the physical world, and robots are one of the most important vehicles for this transition.

What Core Resources Does Google's Robotics Accelerator Provide?

Gemini Robotics Models: Multimodal Capabilities Empowering Robots

One of the core technical resources Google is offering to selected companies is its Gemini Robotics models. As Google's latest-generation multimodal AI model, Gemini possesses powerful visual understanding, language reasoning, and multimodal fusion capabilities. Extending Gemini's capabilities into the robotics domain means robots can more naturally understand human instructions, perceive complex environments, and make sound decisions.

Specifically, Gemini Robotics is a model variant purpose-built for robotics applications, released by Google in March 2025. It comes in two versions: Gemini Robotics and Gemini Robotics-ER (Embodied Reasoning). The core innovation lies in directly connecting Gemini's multimodal understanding capabilities with robot action generation — the model can not only understand visual inputs and language instructions but also directly output control signals at the robot joint level. This "Vision-Language-Action" (VLA) model architecture continues the research trajectory of Google's earlier RT-2 and RT-X series but achieves significant improvements in model scale, reasoning capability, and cross-task generalization. Compared to traditional modular robot systems (with separate perception, planning, and control modules), the end-to-end nature of VLA models dramatically simplifies system integration complexity — a particularly important advantage for resource-constrained startups.

Full AI Technology Stack: Lowering the Technical Barrier for Startups

Beyond the models themselves, Google will also open access to its AI technology stack, including cloud computing infrastructure, TPU compute resources, and robotics development toolchains. For startups, the barrier to accessing these resources is typically extremely high, and the accelerator program significantly reduces the cost of technical validation and product iteration.

TPU (Tensor Processing Unit) is Google's proprietary AI-specific chip, now in its sixth generation (Trillium). Compared to NVIDIA GPUs, TPUs offer unique advantages in large model training and inference efficiency within Google's own frameworks (such as JAX/TensorFlow), particularly excelling in large-batch matrix operations and low-precision inference scenarios. For robotics startups, the compute costs required to train VLA models are extraordinarily high — a single full-scale robot policy training run can require thousands of TPU hours. By opening TPU access through the accelerator program, Google is essentially trading compute subsidies for ecosystem lock-in, a strategy strikingly similar to what AWS and Azure employ with startups in the cloud computing space.

Hands-On Team Support: Deep Technical Collaboration

Throughout the three-month program, Google will assign internal teams to provide hands-on support to the startups. This deep technical collaboration model not only helps startups quickly break through technical bottlenecks but also gives Google a close-up view for identifying and screening potential acquisition or investment targets.

Why Europe as the First Stop?

Google's decision to launch its first robotics accelerator in Europe is driven by multiple strategic considerations:

• Deep manufacturing foundations: Europe boasts a world-leading industrial manufacturing ecosystem. Countries like Germany and Switzerland have deep expertise in precision manufacturing and industrial automation, providing natural deployment scenarios for robotics applications.
• Rich talent pool: European universities maintain world-class research in robotics, control theory, and other foundational disciplines, with institutions like ETH Zurich and Imperial College consistently producing top-tier talent.
• Favorable policy environment: The EU's regulatory framework for AI and robotics is relatively well-defined, providing companies with a predictable development environment.
• Competitive positioning: Against the backdrop of intensifying competition in the robotics space between China and the United States, the European market represents an important third pole, and Google's move also reflects an intent to claim strategic ecosystem territory.

It's worth diving deeper into Europe's academic strengths in robotics. ETH Zurich's Robotic Systems Lab is the birthplace of the quadruped robot ANYmal, and its spinoff company ANYbotics has become a leader in industrial inspection robots. Imperial College London is at the global forefront of dexterous manipulation and soft robotics. Additionally, Germany's Fraunhofer Institute network has long served as a bridge between academic research and industrial application, driving the commercialization of numerous robotics technologies under the Industry 4.0 framework. The Italian Institute of Technology (IIT) is the institution behind the humanoid robot iCub. This complete chain from fundamental research to applied commercialization makes Europe a natural incubation ground for robotics startups and explains why Google chose it as its first stop.

Industry Impact and Future Outlook

The launch of Google's accelerator program sends several important signals:

Large model companies are systematically entering the robotics space. From OpenAI's investment in Figure to Google's launch of Gemini Robotics and its accelerator program, leading AI companies are no longer content competing solely at the software layer — they are actively extending into hardware and the physical world.

Looking back at landmark events in this trend: In early 2024, humanoid robotics company Figure AI completed approximately $675 million in Series B funding, with investors including OpenAI, Microsoft, NVIDIA, and Jeff Bezos. Figure subsequently partnered with OpenAI to integrate GPT-series models into its humanoid robots Figure 01 and Figure 02, enabling the robots to understand task instructions through natural language conversation and execute complex operations. This collaboration was widely regarded as a landmark moment for large model companies entering the physical world. Since then, NVIDIA launched Project GR00T, a foundation model for humanoid robots; Tesla has continued iterating on its Optimus humanoid robot; and Chinese companies like AGIBOT and Unitree Robotics have also secured substantial funding. The entire industry is forming a deep convergence trend of "large models + robotics," and Google's accelerator program is the latest move in this wave.

The "model + ecosystem" competitive model is being replicated in the robotics domain. Just as Android won the mobile internet era through an open ecosystem, Google is attempting to build a Gemini-centric robotics developer ecosystem through its accelerator program.

This analogy deserves deeper examination: During the mobile internet era, Google successfully built a developer ecosystem spanning billions of devices by open-sourcing the Android operating system. Although most hardware profits went to manufacturers like Samsung and Huawei, Google achieved sustained commercial returns through search, app stores, and advertising services. Replicating this strategy in robotics means Google doesn't need to manufacture robot hardware itself — instead, it provides core AI models (analogous to the Android OS) and development toolchains (analogous to Android Studio), enabling numerous robotics companies to build products on the Gemini ecosystem. Once Gemini Robotics becomes the de facto standard in the robotics industry, Google can capture long-term value through model API calls, cloud service consumption, and data feedback loops. This also explains why the accelerator program is offered free to startups — the upfront investment is designed to achieve ecosystem lock-in down the road.

The window of opportunity for robotics startups may be narrowing. As big tech companies begin systematically providing models, compute, and technical support, independent startups that fail to quickly establish differentiated advantages risk being absorbed into these ecosystems.

For the 15 selected startups, this is a rare opportunity to accelerate. But what's even more worth watching is the program's trajectory going forward — whether Google will expand the accelerator to more regions, how Gemini Robotics models actually perform in practice, and whether these startups can ultimately deliver convincing product results. The Physical AI race has only just begun.