In brief
At Generation, we produce detailed research roadmaps on sustainable shifts in industries to identify opportunities for investors and companies to drive positive change. Our new Roadmap Series of articles will highlight our key findings on a different industry each quarter. This first piece focuses on ‘physical world AI’ – robots that can perceive, reason about and act in the real world.
At a glance:
- Over $34 billion of private capital flowed into robotics-related companies in 2025 – more than twice that of 2024.1 Yet some of the best-funded companies are still in the early stages of commercialisation, with scaled deployments years away.
- Physical world foundation models, which include both vision language action and world models, are emerging as the next frontier of artificial intelligence, but data remains a critical bottleneck.
- We see investment opportunities in robotic hardware and the software ‘picks and shovels’ of physical AI, including companies providing data, testing infrastructure and simulation tools.
Why are we focusing on physical world AI?
In 1988, the roboticist Hans Moravec observed something peculiar: tasks that seem effortless to humans, such as catching a ball or folding a towel, were extraordinarily difficult for machines, while tasks that we would consider intellectually demanding, like solving equations or playing chess, were far simpler to automate. This observation became known as Moravec’s Paradox, and for decades it has defined the limits of robotics. We could build machines to play Jeopardy or beat chess grandmasters, but none that could load a dishwasher or safely navigate a city street.
This is now changing. Breakthroughs in computational power, simulation and spatial reasoning models are enabling a new generation of physical world AI. These breakthroughs have captured the attention of investors.
In 2025, over $34 billion of private capital flowed into robotics companies, more than 2.5 times that of 2024.
Companies such as Physical Intelligence, Skild AI and Figure AI raised billions of dollars at multibillion-dollar valuations. Despite early commercial traction, these high valuations reflect the size of the opportunity in robotics: more than half of the $30 trillion US economy is directly tied to physical human labour.2
As sustainable investors, we are excited about the potential for robotics to solve core sustainability objectives across planetary and human health. Here, we define robotics broadly: humanoids and robot arms, self-driving cars, sensors, drones, tractors and more. Many robotics companies are already making significant impact in their respective industries: Zipline is using drones to solve last-mile delivery for medical supplies; Mytra is developing novel warehouse robotics that make supply chains efficient and less wasteful; Carbon Robotics is using robotic weeders to eliminate the use of herbicides; and Bedrock Robotics is building fully autonomous construction equipment to create critical infrastructure faster, cheaper and more safely. At the same time, we are deeply conscious of the potential negative impacts of robotics and autonomy, from prolonging fossil fuel extraction and infrastructure to making weapons fully autonomous. These risks must be tightly managed.
In the last year, leaders in robotics have released product demonstrations that show advancements in simulation, spatial reasoning and dexterity. This slide from our roadmap research shows some impressive examples (images copyright of Figure, Tesla, GitHub and Sunday AI).
Our thesis
Over the past two years, we have met with virtually every type of robotics company: six-axis arm manufacturers (robotic arms with six joints mimicking the movement of a human arm), humanoid developers, simulation companies, autonomous driving labs and foundation labs developing models that help robotics understand and reason about their surroundings.
While robotics hardware companies hold enormous potential, many have yet to demonstrate true reliability and scalability, especially those working on novel ‘form factors’ (the physical design of the robot) like humanoids. Customers are still experimenting with form factors, discovering possible use cases and running pilots that have not yet translated into repeatable revenue. Here, we believe the path from impressive demonstration to commercial deployment remains long and uncertain.
We have been exploring hardware companies taking a more pragmatic approach: using tried-and-tested form factors like six-axis arms, powered by modern software and foundation models, to deliver fast and consistent value to customers. It’s worth noting that robotic arms are hardly new. They have been assembling cars, manufacturing electronics and building solar panels for decades. But these robots have always been specialists: bolted to the floor, programmed to repeat a single motion thousands of times a day and helpless the moment something unexpected happens. Now, with advances in spatial reasoning, we are seeing off-the-shelf robotic hardware become capable of performing more complex, unfamiliar and unstructured tasks, from folding laundry to precisely sorting through piles of mixed recycling and waste.
As the robotics hardware ecosystem matures, we have also focused on companies building at the foundational layer of robotics: the ‘brains’ that will power physical AI regardless of which hardware form factors ultimately scale.
Two types of foundation models are emerging: world models and vision language action models (VLAs). World models, such as Marble from World Labs, simulate the physical environment, predicting how it will evolve over time, giving robots the capacity for spatial reasoning and planning. Vision language action models, such as the π0.6 model from Physical Intelligence, take a different approach, translating visual input and language instructions directly into robotic actions. Most robotics developers are combining both models, using world models for reasoning and planning while VLAs handle real-time execution. We believe that these models will underpin robotics applications in the future, similar to how large language models serve as the foundation for digital AI applications today.
Beyond foundation models, we see significant opportunities to invest in the software infrastructure to build and test modern robotics. Just as companies such as Grafana and Datadog created essential infrastructure for the software industry, hardware requires its own purpose-built platforms for testing and validation, observability and simulation. Some next generation infrastructure providers, including Nominal and Flow Engineering are already being used by sophisticated hardware companies to test prototypes and autonomous systems for safety and functionality in the real world.
Although there are pure-play foundation model companies, such as World Labs, most robotic hardware companies are trying to develop their own world models or vision language action models. This slide from our roadmap shows how we think about the landscape in four buckets.
Three ways we think growth-stage companies can win in the physical AI and robotics market
We believe there are three approaches that growth-stage companies can take to win in this market.
- Leverage a data advantage
Unlike in the digital world, there is no plentiful, open repository of data on which to train and validate autonomous systems. This shortage in robotics data has spurred a race to collect large-scale, high-quality datasets that can help train autonomous systems. In practice, these datasets can include many different types of data. For example, some datasets are videos of humans performing actions that robots would try to emulate. Other data might include readings from sensors that would tell a robot how much pressure to apply when lifting an object. Some companies have developed other creative and distinctive ways of collecting data, including training models on massive, proprietary gaming datasets, gathering data from humans wearing sensor-rich gloves, and employing high-fidelity, sensor-realistic synthetic data that mimic what sensors would perceive in the real world. - Create real value for customers
Our litmus test for whether a company in this space has reached the growth stage is whether it is able to create meaningful, repeatable and reliable value for customers. Although this seems obvious, we believe that there are actually very few companies today in robotics that truly meet the test.
The ability to create sustainable value for customers will distinguish a commercial-ready company from one that is still dominantly a research lab.
There’s another advantage to deploying robots at scale with customers: the more robots you have in the field, the more diverse real-world data you can collect and feed back into your models. - Become essential infrastructure for robotics developers
Across the robotics industry, technologists are engaged in fervent debates about which robotic form factors and model architectures will scale in the long run. But no matter which technology approach ultimately prevails, robotics will need a modern software stack to train and test their autonomous systems.
So far, incumbents like NVIDIA, Texas Instruments and Grafana have provided some of the software tools to support robotics developers, but there's still room for startup companies to perfect solutions to specific, hard problems, like building realistic virtual (“synthetic”) environments to train and test autonomous systems, and monitoring whether robot hardware is performing as expected. We think that some of these tool providers could themselves become enormous businesses, selling to the next generation of robotics companies as well as established manufacturers like ABB, Toyota and BMW. To cite an overused Silicon Valley metaphor, these software companies are selling the picks and shovels of the robotics Gold Rush.
What is the opportunity for investors?
At Generation, we have long been fascinated by the potential of simulation to speed up hardware iteration cycles and unlock new technology in fields ranging from autonomous driving to industrial robotics. Our Growth Equity strategy first started covering simulation and autonomy in 2013 and invested in DeepMap’s Series B investment round in 2018 (DeepMap was later acquired by NVIDIA). And in 2025, we invested in Zeromatter, a company developing simulation infrastructure for large-scale autonomous systems across automotive, aerospace and robotics applications. Simulation is just one example of the picks-and-shovels opportunities we see today in robotics.
One opportunity we are especially excited about is data to test and train robotics systems. Unlike large language models, which could train on the vast amount of human-generated text available online, physical AI requires embodied data – information about how objects move, how forces interact and how to manipulate physical matter. Robotics data simply does not exist at scale, and collecting it is expensive. Every company we spoke with acknowledged this constraint. Some are building teleoperation networks, paying humans to remotely control robots and capturing their movements. Others are investing heavily in simulation, generating synthetic training data in virtual environments. Others have built their own physics and rendering engines to produce synthetic data at an unprecedented scale. We believe that data will continue to be an interesting place to invest as robotics companies scale.
Historically, robotics have been mainly single purpose without the ability to perform complex tasks or reasoning. Now, armed with more powerful chips, advancements in model architecture and improvements in hardware, robotics developers are pushing the boundaries of capability. This slide from our roadmap shows how advances are solving issues (images copyright of Field AI, Generalist and Figure).
Looking ahead
We frequently see announcements and demos from the leading robotics labs on breakthroughs in solving previously intractable problems. We have no doubt that physical world AI can have a real bearing on our economy – from making industrial environments safer to speeding the construction of critical infrastructure like roads, renewable energy assets and factories. We see an exciting pipeline of companies forming for growth-stage investors in this space and are investing in this category.
- Pitchbook: Over $34 billion of private capital flowed into robotics-related companies in 2025 (subscription required)
- Bloomberg: Labor’s Share of US GDP Drops to Record Low in Data Back to 1947 (subscription required)
Important information
The Roadmap Series: How Physical World AI Could Reshape our Economy report is prepared by Generation Investment Management LLP (“Generation”) for discussion purposes only. It reflects the views of Generation as of March 2026. It is not to be reproduced or copied or made available to others without the consent of Generation. The information presented herein is intended to reflect Generation’s present thoughts on sustainable investment and related topics and should not be construed as investment research, advice or the making of any recommendation in respect of any particular company. It is not marketing material or a financial promotion. Certain companies may be referenced as illustrative of a particular field of economic endeavour and will not have been subject to Generation’s investment process. References to any companies must not be construed as a recommendation to buy or sell securities of such companies. To the extent such companies are investments undertaken by Generation, they will form part of a broader portfolio of companies and are discussed solely to be illustrative of Generation’s broader investment thesis. There is no warranty that investment in these companies have been profitable or will be profitable. While the data is from sources Generation believes to be reliable, Generation makes no representation as to the completeness or accuracy of the data. We shall not be responsible for amending, correcting or updating any information or opinions contained herein, and we accept no liability for loss arising from the use of the material.