Bridging the Gap Between Digital Success and Real-World Application

While AI applications like Cursor, Abridge, and Harvey have demonstrated clear value in software-centric domains, this success has yet to be fully replicated in the physical world. Large Language Models (LLMs) leverage datasets from digital sources such as websites, text, and images, enabling them to develop knowledge across multiple subjects. However, physical AI faces a critical challenge: the scarcity of structured, real-world data. Unlike traditional LLMs which process digital information at scale, physical AI models must train on unpredictable, dynamic environmental data, where rules are governed by physics, not code. This disparity has created a significant gap in the ability to understand and interact with the physical world, making scalability far more complex and resource intensive.

Machines Continue to Evolve

Historically, machines have operated within the confines of pre-programmed guidelines and deterministic algorithms, meticulously designed by engineers to respond to specific, anticipated scenarios. This approach ensured precision and reliability in highly controlled environments, such as assembly lines or warehouses, where variability is minimal. However, these systems are inherently fragile when faced with real-world complexity. Every possible situation must be explicitly coded, making them ill-equipped to handle unexpected events or unstructured environments. For example, a robotic arm programmed to pick up objects of a certain size and shape would fail if presented with an unfamiliar item, highlighting the limitations of this rigid, rule-based approach (1).

The advent of machine learning and training-based systems marked a significant leap in the flexibility of physical AI. By learning from datasets, these systems can navigate diverse terrains, manipulate objects of varying shapes and sizes, and detect mechanical failures. For instance, autonomous vehicles trained on millions of driving scenarios are learning to adjust to different road conditions, and robotic arms can adapt to handling new objects with minimal reprogramming. However, this progress comes with hurdles: physical AI still relies heavily on extensive training datasets and often falters when confronted with situations significantly different from those encountered during training. This reliance on data-driven patterns underscores the gap between current capabilities and true adaptability (1).

The next phase of evolution promises to bridge this gap by integrating real-world foundation models that imbue physical AI systems with contextual understanding. These models enable machines to interpret natural language instructions, grasp the broader context of their tasks, and make decisions based on high-level reasoning rather than mere pattern matching. For example, a context-aware robot could understand the instruction "move a shipment to the warehouse, but avoid the restricted area" and adapt its actions accordingly, considering the current state of the environment. This represents a paradigm shift toward general-purpose intelligence, where physical AI systems can execute complex, actionable tasks with minimal human oversight.

Achieving this level of contextual understanding and adaptive reasoning requires a sophisticated technology stack that seamlessly integrates advanced AI software architectures with cutting-edge hardware. Such systems must process vast amounts of real-time data, reason about dynamic environments, and execute actions with precision. This foundation will not only power intelligent physical systems but also redefine their role across industries, from manufacturing and logistics to healthcare and beyond. By moving beyond rigid programming and data-driven patterns, context-based physical AI paves the way for machines that truly collaborate with humans, adapting to the ever-changing demands of the real world.

Bridging the Gap Between Hardware and Software

The hardware tech stack for physical AI is highly specialized, tailored to the demands of specific use cases. For instance, autonomous vehicles rely on LiDAR and sensor arrays for environmental perception, while humanoid robots require servo motors for precise movement coordination. For Physical AI to work, there needs to be an optimal software architecture designed to power real-world AI operations, no matter the hardware components in use:

• Edge Compute and GPUs for efficient model training and inference

• Foundation Models that serve as the backbone for robotics and intelligent systems

• Data Infrastructure to manage and process vast amounts of physical data

• Reinforcement Learning to enable continuous improvement through experiential learning

• Firmware IDEs to connect software with hardware 

• Machine Operating Systems to orchestrate real-time resource allocation

• Comprehensive Security frameworks to protect both digital and physical systems

Each of these components plays a critical role in enabling seamless real-world functionality. Companies like Measurement Labs*, Labyrinth AI, and Mecka are pioneering sophisticated data layers that capture, process, and synthesize information from hardware, laying the groundwork for AI decision-making. Foundational models from World Labs, Physical Intelligence, and Skild are redefining robotics and intelligent systems, while Parallax Labs is pushing boundaries with Reinforcement Learning environments that allow AI to refine expertise through simulated and real-world experiences, ensuring robust performance during complex tasks. Firmware is being transformed by Embedder*, Bootloop, and Rake Labs, whose intelligent coding agents autonomously generate, test, and maintain firmware, reducing human error and accelerating development. Meanwhile, Machine Operating Systems ensure real-time processing and resource management, critical for high-stakes applications. On the security front, Dragos, Claroty, and Spotlight Security provide comprehensive protection frameworks that safeguard systems from cyber threats, ensuring operational reliability across all conditions. Together, these components form a unified ecosystem that bridges the digital and physical worlds, driving the next generation of intelligent, adaptive systems.

*Exceptional portfolio company

Impact Today and in the Future

Live applications of physical AI are already demonstrating tangible business value across industries. In June 2025, Amazon recently deployed its one-millionth robot within one of its 300 facilities with robotic workers. John Deere's autonomous tractors are transforming agriculture by harvesting crops with limited human intervention. Caterpillar is deploying autonomous mining trucks that operate in hazardous environments and improve safety while maintaining consistent productivity levels. These companies, along with many others, are demonstrating how physical AI creates measurable competitive advantages through operational excellence.

The development of physical AI addresses critical labor shortages across manufacturing and industrial sectors through intelligent automation that can augment or even replace existing infrastructure. Physical AI promises substantial cost savings realized through automation while providing unlimited capacity that can scale based on demand. With major market opportunities spanning logistics, construction, agriculture, automotive, and energy, physical AI offers companies the ability to maintain continuous operations with minimal human oversight, potentially revolutionizing industrial productivity and manufacturing economics by transforming labor from a variable cost center into a one-time capital investment with indefinite scalability.

And eventually, the technology will scale to not only automate singular pieces of equipment, but also environments.

Market Opportunity

The convergence of technological breakthroughs, rapid adoption, and strategic investment is propelling the physical AI market into a new era of growth. Advances in computer vision, sensor technology, and machine learning have empowered machines to navigate and operate autonomously in dynamic, real-world environments, unlocking unprecedented capabilities. Industries such as manufacturing and logistics are at the forefront of this transformation, leveraging AI-powered automation to drive measurable productivity gains and reduce human error.

Currently valued at $4B, the physical AI market is projected to skyrocket to $23 billion by 2030, fueled by expanding applications and continued demonstration of ROI. Venture capital is playing a pivotal role in this acceleration, with investments reaching $12B the first half of 2025 (2) – nearly eclipsing the $12.5B total for 2024 (2). This surge in funding underscores investor confidence in the sector’s potential to redefine industries and reshape the future of work.

Conclusion

We are currently in the initial phase of the physical AI evolution, which consists of specialized automation within predictable environments, as demonstrated by Amazon's extensive warehouse robots and John Deere's autonomous tractors. These systems excel in controlled settings but lack the ability to adapt to unpredictable outcomes, such as a tractor halting operations due to unexpected terrain variations, highlighting the limitations of this phase. Phase two will introduce adaptive reasoning capabilities, allowing systems to process verbal instructions and respond to unexpected circumstances, such as a construction robot adjusting its tasks in real-time when faced with sudden weather changes or an emergency response drone improvising a rescue plan in a collapsed building, thereby expanding applications into unpredictable environments such as construction and emergency services. The final phase represents ecosystem-wide coordination, where physical AI evolves beyond individual units to manage complete operational networks, establishing self-governing manufacturing facilities and autonomous supply systems that fundamentally alter economic structures. Physical AI will not simply fill labor gaps but will restructure work itself, eliminating repetitive manual tasks while creating sophisticated roles in system development and human-machine collaboration, benefiting those who embrace cooperative relationships with intelligent technology over those who resist it, with current high investment levels potentially accelerating this progression beyond traditional adoption timelines.

1 Physical AI: Powering the New Age of Industrial Operations

2 Emerging Tech Research Vertical Snapshot: Robotics