Generative Physical Artificial Intelligence & Basic Models for Robots: The Transformation of Robotics through Learning Systems

The end of programming: When machines learn by simply watching – When machines learn to think instead of rigidly obeying.

Robotics is currently undergoing a fundamental paradigm shift that is fundamentally changing how autonomous systems function. While industrial robots have been used in manufacturing for decades, they have thus far been limited to rigid, predefined processes. These machines followed precisely programmed if-then instructions and could only perform the tasks for which they were explicitly coded. Every new requirement, every modified production line, necessitated complex reprogramming by specialized personnel. This traditional robotics was based on deterministic algorithms in which every movement sequence, every gripping position, and every reaction to sensor signals had to be manually defined.

The breakthrough now underway is based on transferring principles known from generative artificial intelligence to the physical world. Just as large language models develop a statistical understanding of language through training on enormous amounts of text, foundation models for robots are now being created that acquire an understanding of the three-dimensional world and physical relationships through observation and simulation. These models are no longer programmed for every single action, but learn generic skills that they can apply to new situations.

Nvidia CEO Jensen Huang calls this moment the ChatGPT moment of robotics, an analogy that underscores the revolutionary dimension of this development. Just as ChatGPT demonstrated to a broad public in November 2022 what modern language models are capable of, Foundation Models could represent a similar threshold for robots. The parallel is not merely metaphorical. The underlying technologies share essential architectural principles. Transformer models, originally developed for language processing, are now being adapted to process sensory data, motion trajectories, and physical interactions.

This development has far-reaching economic implications. The robotics industry is poised for a growth spurt that could dwarf previous developments. While approximately four million industrial robots are currently in use worldwide, market researchers predict that humanoid robots alone could reach twenty million units by 2030. The most ambitious forecasts from ARK Invest anticipate a maximum market volume of twenty-four trillion US dollars for humanoid robots. These figures may seem exaggerated, but they reflect the transformative power that experts attribute to this technology.

Suitable for:

• AI Industry 5.0: How Jeff Bezos' (Amazon) $6.2 billion Project Prometheus is bringing AI to factory floors

From rigid algorithms to adaptive systems

The technological evolution from programmed to learning robots is taking place on several levels. At its core, it involves a shift away from rule-based systems towards data-driven approaches. Traditional robot programming relied on explicit instructions for every eventuality. A robot on an assembly line had to know precisely where a component would be located, its orientation, and the force and speed with which it should grasp it. This precision required structured environments that minimized variability.

Foundation models for robots break with this paradigm by extracting statistical patterns from large datasets. Instead of implementing explicit rules, these models learn implicit representations of tasks, objects, and manipulation strategies. The learning process is similar to human learning through observation and imitation. A model is fed thousands or millions of demonstrations showing how specific tasks are performed. From this data, the neural network extracts patterns and strategies that it can then apply to new, similar situations.

The data for these foundation models comes from various sources. Physical Intelligence collected approximately 10,000 hours of real-world robot data to train its first foundation model. The startup GEN-0 reports an even larger dataset of 270,000 hours of real-world manipulation data from homes, warehouses, and workplaces worldwide. These datasets are enormous, yet they fall far short of the trillions of tokens used to train large language models. The discrepancy is explained by the nature of the data. Robot data is more difficult to collect because it requires physical interactions in the real world. You can't simply download millions of videos from the internet and hope that's enough. Robot data often needs to be actively generated, through teleoperation, human demonstrations, or automated data collection systems.

This is where simulation comes into play, playing a key role in modern robotics research. Physics-based simulators make it possible to generate virtually unlimited amounts of synthetic training data. Nvidia has created platforms like Omniverse and Isaac Sim that provide highly realistic virtual environments in which robots can be trained. The World Foundation Models, which Nvidia is developing under the name Cosmos, generate photorealistic video sequences from simple inputs that respect physical laws and on which robots can learn virtually.

The idea is compelling. Instead of recording millions of hours of real-world interactions, robots can be trained in simulations where time is compressed and thousands of robot instances learn in parallel. The challenge lies in bridging the so-called sim-to-real gap, the discrepancy between simulated and real-world behavior. A robot that performs perfectly in the simulation can fail in the real world if physical properties such as friction, elasticity, or sensor inaccuracies have not been correctly modeled.

The role of German actors in the global robotics landscape

Germany has a long-established robotics industry and is considered one of the leading countries in industrial automation. The robot density in German manufacturing is among the highest worldwide, with approximately three hundred robots per ten thousand employees. This strength in traditional robotics provides a solid foundation, but the question remains whether Germany can successfully manage the transition to cognitive, AI-driven robots.

Several German and European companies are positioning themselves in this emerging market. Agile Robots, headquartered in Munich, has become one of the most ambitious players. In November 2025, the company announced its first humanoid robot, Agile One, specifically designed for industrial environments and slated for production in a new factory in Bavaria by early 2026. Agile Robots emphasizes that the training of its Robot Foundation Model primarily takes place in Munich and is based on real-world production data. A partnership with Deutsche Telekom and Nvidia enables training on the new Industrial AI Cloud, hosted in German data centers and compliant with European data protection standards.

This approach is strategically significant. While many competitors rely on synthetic or generic data, Agile Robots, through its own production and its customers in the automotive and electronics industries, possesses one of the largest industrial datasets in Europe. Data is the lifeblood of artificial intelligence, and access to high-quality, real-world data provides a substantial competitive advantage. The company already has over 20,000 robot solutions in operation and is continuously collecting new data from real-world applications.

NEURA Robotics, based in Metzingen, Germany, pursues a similarly ambitious approach. The company positions itself in the field of cognitive robotics and works closely with Nvidia to develop foundation models for its robotic systems. NEURA emphasizes the combination of real-world data with advanced simulations and has developed a multi-layered AI architecture that combines real-time sensor processing, local inference on the robot, and distributed multi-agent learning. In October 2025, NEURA announced its expansion to Hangzhou, China, with a registered capital of 45 million euros, underscoring the company's global focus.

The German Aerospace Center (DLR) is also investing in foundation models, but with a broader focus on applications in aviation, space, and transportation. The DLR's Foundation Models Adaptation project aims to make large AI models usable for specific applications and to develop lightweight, specialized models. While DLR does not directly develop commercial humanoid robots, its research contributes to the knowledge base upon which industrial players can build.

However, the position of German companies is not without its challenges. Global competition is intense, and both the US and China are investing heavily in robotics and artificial intelligence. In the first half of 2025, China invested six times and the US four times as much capital in AI-enabled robotics as the European Union. This investment gap is worrying. While Europe has invested over twenty billion euros in AI companies, the US is allocating one hundred and twenty billion dollars annually, and China has invested nine hundred and twelve billion dollars in artificial intelligence and related technologies over the past decade.

The regulatory landscape in Europe contributes to this discrepancy. While the AI ​​Act and the GDPR pursue the important goal of promoting responsible AI development and ensuring data privacy, they simultaneously restrict access to training data and increase compliance costs, disproportionately burdening smaller companies. While Europe regulates, US and Chinese companies are experimenting with significantly fewer restrictions.

The economic dimension of technological transformation

The introduction of foundation models in robotics has far-reaching economic implications that extend beyond the robotics industry itself. At its core, it addresses the question of how automation can increase productivity, alleviate the shortage of skilled workers, and secure the competitiveness of highly industrialized economies like Germany.

The training costs for foundation models are substantial and continuously rising. While the original Transformer model cost around nine hundred dollars in 2017, the estimated training costs for OpenAI's GPT-4 were seventy-eight million dollars and for Google's Gemini Ultra one hundred and ninety-one million dollars. These sums far exceed the budgets available to academic institutions or smaller companies. Developing competitive foundation models therefore requires a capital investment that can only be raised by well-funded companies or through government funding.

For robotics-specific foundation models, the exact costs are harder to quantify, but they are likely to be of a similar order of magnitude, if not higher. The need to collect large amounts of real-world robot data requires extensive hardware infrastructure and operational costs. Physical Intelligence reports that its data generation system delivers over ten thousand new hours of robot data weekly. Operating such a system with thousands of data collection devices and robots worldwide is costly.

The return on investment for these projects depends on whether the developed foundation models actually deliver the promised benefits. The economic justification for humanoid robots is based on their ability to replace or supplement human labor in certain areas. A study by Nexery predicts that humanoid robots could automate up to 40 percent of tasks currently performed manually, with a focus on assembly, logistics, and maintenance. The expected payback period is less than 56 hundredths of a year, making humanoid robots an attractive investment.

These calculations are based on the assumption that the acquisition costs for humanoid robots will decrease. While the first models will cost an average of eighty thousand US dollars in 2025, a price of around twenty to thirty thousand dollars is expected by 2030. This cost reduction would be driven by economies of scale, technological improvements, and competition. By comparison, an average industrial worker in Germany costs an employer approximately fifty to seventy thousand euros per year, including social security contributions and benefits. A robot that can work around the clock, requires no breaks, and does not get sick could pay for itself within a few years under these conditions.

The economic impact is ambivalent. On the one hand, automation through cognitive robots could help alleviate the acute shortage of skilled workers in many sectors. Germany and other highly industrialized countries are facing demographic change that is reducing the number of available workers. Robots could fill gaps and maintain productivity. On the other hand, there are concerns that automation will lead to job losses, particularly in sectors involving repetitive, physical tasks.

Historical experience shows, however, that technological progress does not lead to mass unemployment in the long term, but rather to structural shifts in the labor market. New occupational fields emerge that require the maintenance, programming, and monitoring of automated systems. Qualification requirements shift from purely physical labor to technical and cognitive skills. The challenge for education policy is to prepare the workforce for this transformation and to offer retraining programs.

Industry focus: B2B, digitalization (from AI to XR), mechanical engineering, logistics, renewable energies and industry

More about it here:

• Xpert Business Hub

A topic hub with insights and expertise:

• Knowledge platform on the global and regional economy, innovation and industry-specific trends
• Collection of analyses, impulses and background information from our focus areas
• A place for expertise and information on current developments in business and technology
• Topic hub for companies that want to learn about markets, digitalization and industry innovations

The competition for technological leadership

The global competitive landscape in robotics is characterized by a triangle between the USA, China, and Europe, with each region exhibiting distinct strengths and weaknesses. The USA dominates in foundation models for artificial intelligence. OpenAI, Anthropic, Google, and Meta have developed the most powerful language models and possess enormous expertise in scaling neural networks. They are now transferring this competence to robotics. Companies like Figure AI, 1X Technologies, and Physical Intelligence are working intensively on humanoid robots controlled by foundation models.

China has become the world's largest market for industrial robots. In 2024, 54 percent of all newly installed industrial robots were located in China, compared to 17 percent in the European Union. The Chinese government has defined robotics as a strategic priority and is massively promoting the industry through programs such as Made in China 2025. China aims to produce around 40 million robots by 2030, a figure that underscores the government's ambitions. China also leads in AI patents, holding over 70 percent of global generative AI patents, compared to 21 percent from the US and just 2 percent from Europe.

Europe, including Germany, boasts long-established robotics champions like KUKA, ABB, and Stäubli, as well as a strong supplier industry. European strength lies in precision engineering, hardware quality, and a deep understanding of industrial processes. These strengths are valuable, but they are not enough to dominate the field of cognitive robotics. The challenge lies in combining hardware excellence with AI expertise.

The acquisitions and investments of recent years illustrate the shifts in the industry. The takeover of KUKA by the Chinese conglomerate Midea in 2016 was a wake-up call for Europe. SoftBank's recent announcement of its $5 billion acquisition of ABB's robotics division demonstrates that Asian investors are aggressively investing in European robotics expertise. These acquisitions bring capital and market access, but they also carry the risk of losing strategic know-how.

European companies like NEURA Robotics are expanding into China to gain access to this vast market and local resources. While this strategy is understandable from a business perspective, it also raises questions about technological sovereignty. If European robotics companies increasingly relocate their research and development capacities to China, as in the case of Stihl, which moved the development of its robotic lawnmowers there, there is a risk of a long-term loss of expertise.

The answer to these challenges requires a strategic European robotics and AI policy. With its AI Regulation, the EU has created a risk-based regulatory framework that could serve as a global model. However, regulation alone does not create innovation. Significant investments in research, infrastructure, and the training of skilled professionals are essential. The announced partnerships within the EU AI Champions initiative, with over one billion euros in AI investments, are a step in the right direction, but these sums remain modest compared to the US and China.

Suitable for:

• The potential for SMEs-AI-controlled robotics for medium-sized companies: transformation of the world of work and new competitive advantages

Foundation Models as Universal Problem Solvers

The key innovation of Foundation Models lies in their ability to generalize. Traditional robot systems were task-specific, meaning they were tailored to a single task. A welding robot could weld, a gripping robot could grasp, and switching to a new task required complex reprogramming. Foundation Models strive for task generality, the ability to handle a wide variety of tasks with the same model.

This approach is also known as zero-shot or few-shot learning. Zero-shot learning means that a model can solve a new task without specific training for that task by relying on its general understanding. Few-shot learning means that only a few demonstrations are needed to adapt the model for a new task. These capabilities are transformative for robotics because they dramatically increase flexibility.

At CES 2025, Nvidia demonstrated with its Isaac GR00T N1 Foundation Model how a robot can be adapted for new tasks through minimal post-training. The model features a dual architecture inspired by principles of human cognition. System 1 is a fast-thinking action model that enables reflexive reactions. System 2 is a slow-thinking model for deliberate decision-making and planning. This architecture allows the robot to both react quickly to events and handle complex, multi-step tasks.

The company 1X Technologies demonstrated a humanoid robot that autonomously performed household cleaning tasks after being equipped with a policy model based on GR00T N1. The system's autonomy was based on its ability to interpret visual input, understand the context of the task, and execute appropriate actions without requiring every movement to be explicitly programmed.

Franka Emika, a German robotics company, also integrated Nvidia GR00T into its Franka Research 3 system and demonstrated a dual-arm system at Automatica 2025 that autonomously performed complex manipulation tasks. The system was able to infer targets based on camera input and execute appropriate actions in real time, without manual integration or task engineering.

These examples demonstrate that foundation models have the potential to democratize robotics. While programming robots has previously required specialized knowledge, in the future even smaller companies and users without in-depth technical expertise could utilize robots for their purposes. The development of robot-as-a-service models could reinforce this trend by further lowering the barriers to entry.

The importance of data and simulations

The quality of a foundation model depends critically on the data it is trained on. In natural language processing, trillions of words were readily available from the internet, but such vast amounts of data are not easily accessible for robotics. The robot data gap is a fundamental problem. A hypothetical robot GPT, if trained on the same amount of data as a large language model, would require hundreds of thousands of years of data collection, even if thousands of robots were continuously generating data.

Simulations offer a way out of this dilemma. Physics-based simulators can generate virtually unlimited amounts of synthetic data. The challenge lies in ensuring that the behaviors learned in the simulation are transferable to the real world. Various techniques are used to bridge the sim-to-real gap. Domain randomization systematically varies physical parameters in the simulation, making the model more robust against real-world variations. Reinforcement learning with human feedback allows models to be trained using reward signals derived from both simulations and real-world interactions.

Nvidia Cosmos, designed as a World Foundation Model, generates photorealistic video sequences from simple inputs, serving as training environments for robots. The idea is that robots can learn in these generated worlds without the costs and risks of real-world experiments. The model understands physical properties and spatial relationships, ensuring that the generated scenarios are realistic.

Another promising approach is the use of human video data. People perform millions of manipulation tasks daily, which are recorded on video. If it becomes possible to extract relevant information for robot learning from these videos, the database could be significantly expanded. Vision-language models like CLIP have shown that visual concepts can be learned from natural language, and similar approaches are now being explored for robotics.

German and European research institutions are contributing to these developments. The Fraunhofer Institute for Material Flow and Logistics is working on robotic simulations and machine learning systems. The German Research Center for Artificial Intelligence (DFKI) is developing AI methods for robot learning. This research is fundamental to the competitiveness of European companies, but it must be supported by sufficient funding and the transfer of knowledge into industrial applications.

Challenges and open questions

Despite the enormous progress, numerous challenges remain. The robustness of foundation models is a key issue. A model that performs well in a test environment may fail in the real world when faced with unexpected situations. The generalizability, touted as a major advantage, must prove itself across a wide range of scenarios.

The safety of autonomous systems is another critical dimension. As robots increasingly operate autonomously and make decisions based on foundation models, how can it be guaranteed that they behave safely and do not endanger humans? Traditional robotics relied on hard-coded safety mechanisms. With learning systems, such strict boundaries are more difficult to implement.

The ethical and societal implications of cognitive robotics are being intensely debated. The question of responsibility is being redefined. If a robot makes a decision that results in harm, who bears the responsibility? The robot's manufacturer, the developer of the foundation model, the operator, or the robot itself? These questions are not trivial and require legal and regulatory clarification.

The impact on the labor market is a subject of much debate. While some experts argue that robots will alleviate the skills shortage and create new jobs, others fear that low-skilled workers in particular could be displaced. One study estimates that humanoid robots could automate up to 40 percent of manual tasks. The societal challenge lies in managing the transition in a way that ensures the benefits of automation are distributed fairly and social disruption is minimized.

The strategic importance for Germany and Europe

The development of cognitive robotics is not only a technological but also a geopolitical issue. The ability to develop and produce intelligent robots is increasingly seen as a strategic factor. Robotics is finding applications not only in civilian sectors but also in defense, where autonomous systems are gaining importance.

Germany has the potential to take a leading role in cognitive robotics if the right framework is established. Its strengths lie in precision mechanics, software development, and a deep understanding of industrial processes. The automotive industry, historically a key driver of robotics, could once again play a central role. Its established supplier networks and extensive data pool from millions of real-world manufacturing processes are valuable assets.

However, this potential must be actively harnessed. A robotics strategy for Germany and Europe should encompass several elements. First, significant investments in research and development are needed to keep pace with the US and China. Second, the regulatory framework must be designed to foster innovation rather than hinder it, without compromising safety and ethical standards. Third, collaboration between industry, research institutions, and startups should be intensified to accelerate the transfer of knowledge into marketable products.

Promoting entrepreneurship and creating an attractive environment for robotics startups are crucial. Many of the most innovative developments come from agile and risk-tolerant startups. Germany and Europe must ensure that such companies have access to capital, talent, and markets.

The training of skilled workers is another critical factor. The demand for experts in artificial intelligence, robotics, and related fields far exceeds the supply. Universities and vocational schools must adapt their curricula and increase training in these areas. At the same time, retraining programs should be offered to existing workers so they can manage the transition to an automated workforce.

From rigid machines to learning partners – Europe's path into the robotics era

The transformation from programmed to learning robots represents one of the most significant technological shifts of the coming decades. Foundation models for robots have the potential to dramatically expand the flexibility and application possibilities of autonomous systems. Robots will no longer be rigid machines that only perform predefined tasks, but adaptive systems that can learn from experience and adjust to new situations.

The economic implications are far-reaching. Automation through cognitive robots could increase productivity in numerous industries, counteract the skills shortage, and strengthen the competitiveness of highly industrialized economies. Market forecasts point to exponential growth, with the potential for trillions of dollars in added value.

Germany and Europe face the challenge of combining their traditional strengths in robotics with the new demands of cognitive systems. The hardware excellence of German and European companies provides a solid foundation, but it must be complemented by AI expertise. Companies like Agile Robots and NEURA Robotics demonstrate that European players are indeed capable of competing in this field. However, global competition is intense, and both the USA and China are investing heavily in this future technology.

This development requires a systemic approach that involves research, industry, politics, and society. Technological innovation must be accompanied by smart regulation that ensures safety and ethical standards without stifling innovation. The societal debate about the impact of automation must be conducted constructively to alleviate fears and highlight the benefits.

The transition from programmed to learning robots is more than just technological progress. It marks the beginning of a new era in which machines are no longer mere tools, but partners who work alongside humans to tackle complex tasks. How societies shape this transition will determine whether the benefits of this technology are widely shared and whether Europe can play a leading role in this new world. The opportunities are enormous, but they must be seized. The time to act is now.

Here you will learn how your company can implement customized AI solutions quickly, securely, and without high entry barriers.

A Managed AI Platform is your all-round, worry-free package for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a turnkey solution tailored to your needs from a specialized partner – often within a few days.

The key benefits at a glance:

⚡ Fast implementation: From idea to operational application in days, not months. We deliver practical solutions that create immediate value.

🔒 Maximum data security: Your sensitive data remains with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We handle the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & Scalable: Your AI grows with you. We ensure ongoing optimization and scalability, and flexibly adapt the models to new requirements.

More about it here:

• The Managed AI Solution - Industrial AI Services: The key to competitiveness in the services, industrial and mechanical engineering sectors

☑️ Our business language is English or German

☑️ NEW: Correspondence in your national language!

 

I would be happy to serve you and my team as a personal advisor.

You can contact me by filling out the contact form or simply call me on +49 89 89 674 804 (Munich) . My email address is: wolfenstein ∂ xpert.digital

I'm looking forward to our joint project.

 

 

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitalization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

Xpert.Digital has in-depth knowledge of various industries. This allows us to develop tailor-made strategies that are tailored precisely to the requirements and challenges of your specific market segment. By continually analyzing market trends and following industry developments, we can act with foresight and offer innovative solutions. Through the combination of experience and knowledge, we generate added value and give our customers a decisive competitive advantage.

More about it here:

• Use the 5x expertise of Xpert.Digital in one package - starting at just €500/month