The Physical AI & Robotics Landscape

By Lila Paul and Hanna Edgren

The Next Frontier of Artificial Intelligence

In recent years, artificial intelligence has developed rapidly, with investing activity moving from generative AI to agentic AI and towards physical AI. Three big shifts have characterized its development: from prediction to generation, from text to reasoning across a variety of sources, and from passive tools to systems that can make decisions and act upon them. 

Artificial intelligence has moved beyond software development into the physical world, where it can perceive, understand, and interact with real world environments. Simultaneously, the world of robotics is undergoing its own transformation from preprogrammed machines doing simple tasks to increasingly autonomous systems powered by AI. Together, both these developments introduce a broad, emerging, and fast-moving market of physical AI. 

As AI begins to move from digital outputs to physical action, the next question is what distinguishes physical AI from earlier forms of robotics.

Defining Physical AI

Physical AI refers to AI systems that can see, understand, and make actions with the real world. Instead of text-based outputs, the software tells the system to perform an action. These systems adapt to their surroundings as they learn and become more advanced. Common examples are cameras, sensors, robotic surgery arms, and autonomous vehicles. So far, humanoids are the most advanced expression of physical AI. But physical AI is not just important because it represents a technical innovation, physical AI solves real-world problems across a variety of industries.

The Technical Inflection Point

There are four main areas where technology has improved in the physical AI space: agentic intelligence, simulation & synthetic data, edge compute, and hardware. 

Recent developments in AI stimulated the development of agentic intelligence, which allows AI systems to make decisions in unfamiliar situations. Agentic AI can read and reason and interpret information given from user prompts, act by calling on specific tools, and then reflect and evaluate the outcome until the goal is met, often turning complex goals into many steps. It also enables adaptation and replanning in changing environments, which is particularly applicable in the real-world. 

In addition to advances in AI models, edge computing and robotics hardware have also improved. Smaller, optimized models can run closer to the device which reduces reliance on cloud and data-center compute. This modification leads to lower latency for actions which robots need in complex environments. Simultaneously, advancements in sensors, actuators, batteries, and GPUs are making robotic systems more reliable and cost-effective.

Although AI can adapt to new situations, AI still needs intense training. Real data is hard to obtain and it is difficult to collect in large amounts. Sim2Real combats this issue, by creating a virtual world that trains physical AI and Real2Sim improves these simulations with real-world failures. These simulations also allow the physical AI to train for rare edge cases as the world is often quite unpredictable. 

Demand and technological innovation converging has created a robust, fast-moving, and lucrative physical AI market.

This market can be segmented into seven key categories and mini-markets, representing the full Physical AI Stack (see Figure 3).

Compute & Silicon refer to edge AI chips, memory, cloud compute, and training servers, while sensing and perception encompass LiDAR¹, cameras, microphones, and radars. Data & Simulation together with Foundational Models & Training combine all types of data used to train different AI models. Hardware Platforms focus on the actual body of the system, the components used to make it. Finally, Integration & Middleware includes management software, operating systems, and coordination layers that enable the final category, Applications & Deployments, where robots and physical AI systems are applied in real-world environments.

Although there are many opportunities, physical AI faces many challenges on its road to full adoption.

Market Map 

The physical AI landscape has been organized into three main categories: foundation models and training, hardware platforms, and applications and deployment. The market map shown below specifically focuses on applications in healthcare and the life sciences.

Bottlenecks to Adoption

Unlike software, physical AI interacts in the real world. Therefore, it is harder to implement and deploy. A few prominent bottlenecks emerge that block this path to adoption. The real-world is unpredictable and AI often struggles when it sees something it was never trained on. The real-world data that the AI needs and the structural components of the robotic instrument are often expensive and difficult to collect. Additionally, the memory needed to store all the training and data that it does have is immense. Finally, malfunctions and AI hallucinations have real-world impacts as the robots interact with the world people live in. These bottlenecks show why the physical AI market is forming across all layers in the physical AI stack. Companies are not only building robots, they are building the foundational models, hardware, and applications needed to make physical AI feasible in the real world.

Many believe that with the improvements in AI, it will begin to replace jobs. 53 percent of Americans fear AI will take their jobs, which is a real but overexaggerated fear. Every generation faced this fear with the invention of new technology, whether it be deployment of machines in the Industrial Revolution or computers in the 20th century. Some jobs were eliminated but people adapted and many of these events actually created jobs. Even with recent developments in AI, US unemployment remains at low at 4.3% as of June 2026. Jobs will disappear but predictions and fears are much higher than anticipated. However, the jobs that AI targets are more junior roles and ones that can be easily automated. Jobs in markets with high transaction frequencies are easier and more likely to become automated by AI rather than jobs. These include, equity options, stocks, crypto, and FX. 

BOV Approach

By leveraging AI, infrastructure developments, and the increasing commoditization of robotics hardware, new companies are addressing many of the problems the physical AI and robotics industry face. Instead of building specific robots, companies are combining multiple parts of the AI stack to support a variety of tasks. 

The direction of physical AI is clear. As AI systems become more advanced and hardware more deployable, AI will move into the physical world. Although many systems are already deployed such as safety and warehouse picking systems, they are gaining more traction and becoming more and more common as technology advances. The companies that will succeed will be those that solve labor, safety, efficiency, and cost problems with reliable and capable systems.

Portfolio company, Conceivable Life Sciences, is leveraging robotics to improve outcomes and access to IVF. Conceivable’s, AURA, is physical AI being brought to the IVF lab – a system that perceives, reasons, and executes the 200+ precise steps it takes to turn an egg and sperm into an embryo. Another portfolio company, Oath Surgical, is leveraging robotics and technology to deliver improved surgical care and outcomes in the outpatient setting. Robotics coupled with OathOS create clinical, operational, biological, and spatial data that will be immensely valuable for the future of physical AI.

We are excited by companies that are building infrastructure and vertical applications in healthcare, life sciences, and security. Please reach out!