Physical AI: When machines start to think and act in the real world
Redhat.com·July 21, 2026
AI Summary
Physical AI represents a shift beyond traditional automation, where intelligent systems actively solve real-world problems in operations, security, and maintenance. This emerging technology enables machines to think and act autonomously in physical environments rather than being limited to scripted tasks.
If you read the first part of this article, you already know that physical AI isn't just another wave of automation. You saw how intelligent systems are solving real problems in operations, security, and maintenance in ways that traditional robotics simply can't. And you were introduced to the loop that makes it all work: perceive, reason, act.
In this second part, we go deeper. We look at why AI is such a transformative ingredient for robotics specifically, why edge computing isn't an optional add-on but a hard requirement, and how Red Hat helps organizations build the open, enterprise-grade foundation that makes physical AI possible at scale, without trading one form of vendor dependency for another.
Physical AI refers to AI systems that operate in and interact with the physical world. It's the combination of robotics, AI, and edge computing working together so machines can sense what's happening around them, reason about it, and take action, all in real time. Unlike traditional robots, which follow fixed programmed sequences and break when conditions change, physical AI systems learn, adapt, and improve as they operate.
The 3 challenges we explored in Part 1—operations automation, security and surveillance, and maintenance and inspection—all share a common thread. They require systems that can handle unpredictability, work continuously without human supervision, and get smarter over time. Traditional automation was never built for this. physical AI is.
The perceive-reason-act loop introduced in Part 1 is useful as a description of how physical AI works, andas a map of what it requires underneath.
Perceiving the physical world in real time requires the right hardware at the right place: sensors, accelerators, and edge infrastructure that can process data locally without cloud dependency. Reasoning about that data requires AI models sophisticated enough to interpret complex sensor inputs, simulate possible futures, and make decisions in milliseconds. Acting on those decisions requires robust robotics platforms, control systems, and the middleware that ties everything together.
To understand why AI in particular is such a powerful ingredient here, it helps to look at what happened in other fields when AI replaced rule-based approaches.
Before large language models (LLMs), natural language processing (NLP) relied on handcrafted rules, grammars, and dictionaries. These systems worked only in narrow conditions and broke the moment they encountered something they weren't explicitly programmed to handle. Then deep learning (DL) changed everything—suddenly systems could understand context, handle variation, and improve through experience. The same transformation happened in gaming, where systems like AlphaGo replaced manually encoded rules with learned strategies that eventually exceeded human performance. And in computer vision, where rule-based filters and edge detection pipelines gave way to AI models that can recognize objects in any lighting condition, from any angle, in real time.
Full story reconstructed from Redhat.com. Formatting and media may differ from the original.
Adaptability. Traditional robots break when conditions change, even slightly. AI-driven robots learn from their environment and keep working. A robot trained to pick one type of object can, with AI, start handling variations in size, orientation, or packaging it has never seen before, without anyone reprogramming it.
Continuous learning. A traditional robot does exactly what it was programmed to do, and nothing more. AI-driven systems improve through experience. Simulation-based training combined with real-world operation creates continuous learning loops where the longer the system runs, the better it gets.
Generalization. Rule-based systems fail outside the exact conditions they were designed for. AI models apply what they have learned to situations they have never encountered. Companies are already building generalized robot brains that enable any robot to master new tasks with minimal retraining, across different robot types and environments.
Natural language and human collaboration. With vision-language-action (VLA) models, a robot can receive instructions in plain language and act on them without explicit programming. A warehouse robot can respond to, "organize the returned items by category" without a specialist configuring every product type. VLA models are enabling systems to perceive, reason, and act with increasing autonomy alongside humans in environments that are far less structured.
Sensor fusion and real-time perception. Traditional robots see the world through one sensor type. AI-driven robots process multiple streams simultaneously (camera, LiDAR, thermal, tactile) and build a coherent picture of the environment in real time. This is what enables a drone to map terrain while avoiding obstacles, or a cobot (collaborative robot) to work safely next to a human.
Predictive reasoning and world modelling. AI gives robots the ability to simulate possible futures before acting. Rather than reacting to what's happening now, systems equipped with world models can predict what's about to happen and plan accordingly. This is what makes predictive maintenance possible and what allows a robotic arm to plan a complex movement path without trial and error in the real world.
Collective learning at scale. Traditional software deployment is a one-way channel: human logic flows into machines, and that is where the knowledge stops. An AI-enabled fleet amplifies this. Every robot that encounters an unfamiliar object, an edge-case failure, or an unexpected environment contributes that experience back as training data. The model improves from the aggregate of everything the fleet has collectively seen, and then redeploys that improvement to every machine at once. Combined with the generalization capability AI brings, this shared learning loop is what makes the fleet scale smarter and faster than any single robot or human engineering team could achieve independently.
This isn't something that might happen in the future, it's already happening in factories, warehouses, facilities, and infrastructure around the world.
All of these capabilities depend on one thing that is easy to overlook, however—the AI has to run somewhere. And for physical AI, that somewhere can't be a distant cloud. This brings us to the second pillar.
Edge computing makes physical AI real. While physical AI defines what the machine does, edge computing determines where the AI computes: locally, on the device, rather than in a distant data centre. For physical systems that interact with the real world in real time, this is a fundamental requirement.
Zero tolerance for latency. When a robotic arm needs to catch a falling object, it can't afford the round-trip delay of sending sensor data to a cloud and waiting for a response. Edge computing runs AI inference locally, enabling millisecond-level responsiveness.
Uncompromised reliability. Physical machines in industrial environments, in remote locations, or on moving vehicles, can't simply pause when the internet connection drops. Edge-deployed models mean that the sense-reason-act loop can continue regardless of network status.
Massive bandwidth efficiency. Physical AI systems generate enormous amounts of continuous sensor data: multiple high-resolution camera streams, LiDAR point clouds, audio feeds. Sending all of this to a cloud is expensive and slow. Local edge processors handle the sensor fusion and transmit only the relevant metadata back to central systems.
Privacy and security. Processing sensitive environmental and operational data strictly on-device limits exposure to external network risks, which matters a great deal in industrial, healthcare, and security contexts.
Moving AI to the edge isn't simple. It requires hardware powerful enough to run inference in real time, but the hardware is only the starting point. The software stack underneath needs to deliver security policies that isolate models from each other and from the underlying system, through mechanisms like SELinux, cgroups, and containers. On top of that, an orchestration platform needs to manage workload lifecycles across the entire fleet at scale. And all of this needs enterprise-grade support and certification.
This is why Red Hat is working with partners to bring enterprise-grade infrastructure to edge hardware designed for physical AI, such as NVIDIA, with the NVIDIA Jetson family support on Red Hat Device Edge, or with Intel and the Intel Ultra Core 3. It's a concrete example of what it looks like when the software foundation and the physical hardware are built to work together from the start.
As mentioned earlier, physical AI sits at the intersection of 3 technology domains: edge computing, AI, and robotics. Each is complex on its own and the integration between them is where things get difficult, and where platform choices have long-term consequences.
Red Hat's is providing the open, enterprise-grade foundation that makes this integration manageable at scale, without locking organizations into any single vendor's roadmap.
Red Hat OpenShift AI manages the full AI model lifecycle, from training through production deployment and monitoring, across hybrid cloud and edge environments. For physical AI, this means that you'll have a complete end-to-end MLOps lifecycle for the computer vision, predictive maintenance pipelines, and VLA models running on edge devices.
One of the most relevant capabilities for edge and physical AI deployments is model optimization. Edge hardware has real constraints: limited memory, limited power, and limited compute. Red Hat's acquisition of Neural Magic in early 2025 brought deep expertise in inference performance engineering and model optimization directly into the Red Hat AI portfolio.
Additionally, our collaboration with NVIDIA helps make enterprise AI practical with validated support for NVIDIA GPUs and Jetson edge devices, integration with NVIDIA NIM for generative AI deployment, and the kind of enterprise support that research-grade tools do not provide.
Red Hat Enterprise Linux (RHEL) brings capabilities that matter specifically for physical systems at the edge. The real-time kernel delivers the deterministic, low-jitter performance that robotic coordination, motion control, and machine vision require at the edge, without the proprietary lock-in of traditional real-time operating systems. Image mode for RHEL keeps fleets consistent through atomic, immutable OS updates that either complete cleanly or roll back, with no partial states that could leave a physical system in an undefined condition. For smaller devices, Red Hat Device Edge with MicroShift brings Kubernetes to hardware running on as little as 2 CPU cores and 2 GB of RAM, extending the same APIs, security posture, and operational tooling from the data centre to a robot or drone on a factory floor. Fleet management across all of this is handled centrally through Red Hat Edge Manager.
Red Hat's software and hardware certification programs cover applications, servers, accelerators, and partners across the ecosystem, and Extended Life Cycle support provides up to 14-year lifecycles (and even more than that) with 24x7 service level agreements (SLAs) and security patches delivered before public disclosure. For physical infrastructure that runs continuously and can't be replaced on a software vendor's schedule, these commitments matter.
Red Hat contributes to Open Source Robotics Alliance (gold member) and the Robot Operating System (ROS), which means early visibility into vulnerabilities and influence over the direction of foundational technologies. Fedora Robotics SIG is actively working on ROS, and because Fedora is where Red Hat stabilizes packages before they move into RHEL, this work is creating a clear path for ROS workloads to run on enterprise-grade, fully supported infrastructure, something the robotics community has long needed.
Red Hat also validates support for x86 and ARM architectures, with RISC-V currently in technical preview . This hardware breadth, combined with a certified partner network that provides interoperability across the full stack, means organizations aren't building on a platform that will become a dead end.
And all of this is available while also avoiding vendor lock-in. This is one of the most important conversations to have before choosing a physical AI platform. Legacy automation systems have created expensive dead ends for many organizations because proprietary vendors control the roadmap, the integrations, and the pricing. Red Hat's open source foundation, its upstream first philosophy, and its contribution to the Linux kernel, glibc, KVM, GCC, and robotics middleware all mean that the technologies Red Hat builds on are governed by the broader community, not by a single vendor. Organizations choosing Red Hat are choosing a platform that evolves with industry needs rather than one that constrains them.
The most common mistake is trying to do too much at once. We suggest a more measured approach.
Choose a first use case that is bounded (a single location or process), measurable, not too complex (but also not too simple), visible enough that a successful result will be noticed across the organization, and where related data can be collected, curated, and annotated.
Choose a platform that can adapt and scale with future needs, and evaluate the risk of vendor lock-in from the start.
Build the foundation: deploy managed edge nodes on target hardware, and integrate with the enterprise systems (WMS, MES, ERP) the operation already depends on.
Finally, validate in simulation before going live in the real world. Define success criteria. Measure results. Scale. Then select your next use case.
Physical AI isn't a single project, it's a capability that compounds. The organizations that take the first step thoughtfully, with the right platform and the right partners, are those that will be leading their industries as this technology matures.
Senior Principal Specialist Solution Architect
Luis Arizmendi is a tech enthusiast at the intersection of Edge Computing, Artificial Intelligence, MLOps, and Physical AI, with a deep background in containerized workloads at scale, and infrastructure optimization for high performance or real-time deterministic operation.
At Red Hat, he helps organizations design scalable architectures and bring AI beyond the data center, all the way to the edge, into systems that must perceive, decide, and act in the real world.
Passionate about open source technology and the convergence of AI, Edge, and Robotics into the next wave of intelligent, autonomous systems.
The latest on IT automation for tech, teams, and environments
Updates on the platforms that free customers to run AI workloads anywhere
Explore how we build a more flexible future with hybrid cloud
The latest on how we reduce risks across environments and technologies
Updates on the platforms that simplify operations at the edge
The latest on the world’s leading enterprise Linux platform
Inside our solutions to the toughest application challenges
The future of enterprise virtualization for your workloads on-premise or across clouds
