NVIDIA, Google DeepMind, and Intel Make Successive Moves as Physical AI Competition Shifts to Infrastructure

Recently, physical AI is becoming a new hot spot of common interest to both the technology industry and the capital markets. In particular, at a time when some technology sectors such as semiconductors and storage have experienced market volatility, physical AI-related concepts have nonetheless shown a certain degree of heat. Behind this is a reflection of the market's expectations for the next stage of AI development: as large models gradually move from the digital world into the real world, AI that can understand the physical environment, interact with real devices, and autonomously execute tasks may open up new industrial space.

In late October 2026, three noteworthy developments emerged in succession in the field of physical AI:

On October 21, NVIDIA announced at the GTC Berlin conference an expansion of the Isaac open robotics development platform, adding a humanoid robot reference platform, datasets, and a runtime, and updating its robot learning framework.

On October 27, Google DeepMind released research results on the Genie 4 world model, improving the realism of interactive simulated environments and adding native audio generation capabilities, enabling users to change simulated scenes in real time through natural language.

At the same time, Intel and SAP demonstrated a proof of concept (PoC) for physical AI and autonomous warehouses, converting business tasks in an enterprise warehouse management system into operations that robots can execute, and feeding execution status and results back to the enterprise system.

The three developments come respectively from world models, robotics development platforms, and industrial applications, yet they jointly point to a key shift in the development of physical AI: the focus of competition is extending from merely improving the intelligence level of models to building a complete set of technological and industrial infrastructure—enabling AI to understand the physical world, learn in simulated environments how to act, and ultimately complete tasks through real devices and create measurable business value.

Traditional generative AI mainly processes text, image, and video information in the digital world, whereas physical AI faces a more complex real-world environment. It not only needs to recognize objects, but also to understand the spatial relationships, shapes, and attributes among objects, judge how they move, and determine what outcomes different actions may trigger. For a robot, seeing a cup is only the first step: it also needs to determine the cup's position, shape, and grasping method, and predict the force changes that may occur during grasping.

This means that physical AI cannot simply generate actions directly based on current perception results; it also needs the ability to predict future states. World models are providing a new technical approach for this: by building internal models that can represent environmental states and simulate environmental changes, AI can, before actually taking action, first simulate in a virtual environment the possible outcomes of different actions, and then adjust its decisions accordingly.

On October 27, Google announced the release of Genie 4, its most realistic and most interactive world model to date. Unlike traditional video generation models that mainly generate video clips based on prompts, Genie 4 can build explorable virtual environments based on text instructions and adjust scenes in real time as users continue to input instructions, allowing users not only to watch AI-generated worlds but also to enter them and interact with the environment. For example, users can ask the scene to switch from sunny to heavy rain, change the camera perspective, or add new characters and objects to the original environment. The system executes these instructions while the simulation continues running, rather than generating a separate video from scratch each time.

Notably, Google is also trying to combine Gemini's reasoning capabilities with Genie 4's world simulation capabilities. In this architecture, Gemini is responsible for observing the simulated environment, understanding task goals, breaking complex tasks into multiple steps, and continuously checking execution progress; Genie 4 generates and adjusts simulated scenes in real time according to these instructions. In an experiment demonstrated by Google, Gemini 4 Argon broke down a user-proposed task into a series of steps, then passed these steps as real-time instructions to Genie 4, simulating Apptronik's Apollo humanoid robot completing a chess-playing task.

Of course, the potential applications of world models are not limited to robots. Fields such as game development, interactive education, and film and television production can all use them to generate explorable and adjustable virtual environments. But for physical AI, the more important value lies in the fact that they are expected to provide richer training and testing environments for robots and agents, allowing them to simulate different tasks and complex scenarios before entering the real world.

For humanoid robots, world models have especially important application potential. Currently common vision-language-action models (VLA) can directly generate robot actions by combining camera images and human instructions. For familiar tasks, this approach has already made notable progress. But in complex scenarios such as grasping fragile objects, robots also need to judge contact, forces, collisions, and object motion more accurately. World models can work in coordination with VLA to first understand the current environment and task, rehearse the possible outcomes of different operations, then select a more appropriate action and adjust based on actual feedback after execution. This helps improve the success rate and exception-handling capability of humanoid robots in unfamiliar scenarios.

Building a full-stack system from data to deployment

If world models address how AI understands and simulates the physical world, then robotics development platforms address how to train these intelligent capabilities and ultimately deploy them onto real robots. From data collection and model training to simulation testing, edge computing, and on-site operation, for a robot to truly possess autonomous action capabilities, it needs a complete set of interconnected technical infrastructure.

Take NVIDIA's expansion of the Isaac open robotics development platform this time as an example; its layout already covers multiple key stages of robot development—

Data infrastructure: Human 1M includes diverse human activities, synchronized three-view video, and 3D hand-tracking data; SimReady Central Library provides more than 2,000 simulation-ready assets, offering data and environmental resources for robot training.

Simulation training: Isaac Lab 3.0 is an open-source learning framework that supports agents, using multi-GPU workflows and physics engines including Newton and NVIDIA PhysX to support training and optimizing robot policies from desktop to data center.

Runtime and deployment: Isaac OS is an open-source runtime framework built on NVIDIA Holoscan that seamlessly connects AI, sensors, control systems, and robot hardware, thereby achieving deterministic real-time performance from simulation environments to physical robots.

Edge computing and security: Computing platforms such as Jetson Thor provide robots with local inference and real-time control capabilities, while Halos further focuses on security issues during robot deployment.

From data resources and simulation training frameworks to runtime software, edge computing platforms, and security systems, NVIDIA is integrating the previously relatively fragmented technical stages of robot development. This also shows that NVIDIA is not content to merely play the role of a robot AI chip supplier; rather, it hopes to build a technology ecosystem covering the entire lifecycle of robot development through computing power, models, data, simulation tools, and software frameworks.

This brings to mind the path by which NVIDIA built its GPU software ecosystem through CUDA—CUDA not only made it easier for developers to use GPUs for parallel computing, but also gradually became important infrastructure in the field of AI computing. Today, NVIDIA hopes to extend similar ecosystem advantages into robotics through Isaac: developers use its tools to generate data, train models, conduct simulations, and further deploy applications on platforms based on its computing power. As more robotics companies and developers work around this system, the value of Isaac will no longer be limited to a particular piece of software or hardware, but may be reflected in the clustering effect of the entire development ecosystem.

A robot having perception, reasoning, and action capabilities does not mean it can already truly integrate into enterprise production and operations. Physical AI also needs to understand enterprise business goals and translate tasks in business systems into execution activities in the real world. The autonomous warehouse proof of concept (PoC) demonstrated by Intel and SAP provides a good case for observation.

This solution combines the business scenario information and workflow orchestration capabilities provided by the SAP Extended Warehouse Management system (SAP EWM) with edge AI and robotics software and hardware based on Intel technology, demonstrating how physical AI can automate warehouse material handling processes and translate business requests in enterprise systems into execution tasks for autonomous robots. Specifically, task requests are initiated by SAP EWM and dispatched to autonomous robots through Intel's edge AI hardware and software. The robots are responsible for transporting materials and feeding task status and completion back to SAP EWM, thereby establishing a closed loop between enterprise workflows and physical execution. This PoC demonstrated improvements in operational efficiency, including shortening material transport distances and reducing repetitive manual handling work.

Intel and SAP demonstrated the practical application of physical AI through a live robot demonstration at the SAP TechEd conference

Intel provided the technical foundation for edge AI and robotics, including scalable heterogeneous computing capabilities, that is, through CPUs, GPUs, and NPUs working together to complete environmental perception, reasoning, path navigation, detection, and real-time control locally. Around physical AI, Intel has also built a technology system covering industrial-grade chips, real-time computing capabilities, and open software, including OpenVINO Physical AI, Physical AI Studio, and Robotics AI Suite, aimed at reducing the integration complexity of different software and hardware systems and supporting solutions in moving from proof of concept to scalable, repeatable deployment.

Seeing this case, some people may wonder—"Isn't this just an ordinary robot material handling scenario?"

In the author's view, the core difference lies in the fact that traditional robotic automation tends to revolve around pre-set tasks and relatively fixed control flows, whereas this solution attempts to further connect enterprise business systems with physical execution systems. SAP does not merely send a transport instruction to the robot; it provides the business context, workflows, and process rules behind the task. Intel's edge AI technology, in turn, provides computing support for the robot to perceive the environment, assess conditions, and execute tasks on site. More importantly, the status, results, and exception information during robot execution can be sent back to SAP EWM, enabling the enterprise business system not only to issue tasks but also to understand the actual execution of those tasks.

This means the robot is no longer merely a relatively independent automation device, but is incorporated into the enterprise's overall business processes. Enterprises can schedule physical devices based on business needs and understand on-site execution through the information fed back by the devices. This two-way collaboration between IT and OT also provides a reference for expanding physical AI to different factory sites, robot platforms, and operational scenarios in the future.

From this, it can be seen that the key to physical AI truly entering industrial application is not merely enabling robots to connect to AI models and gain stronger perception and action capabilities, but rather integrating intelligent capabilities into enterprise business processes and establishing a complete closed loop from business decision-making to task execution to result feedback. Only when AI can not only understand the real world but also understand what an enterprise wants to happen in the real world and drive it to happen through physical devices does physical AI truly begin to create measurable business value.

Conclusion

From Google's world model to NVIDIA's robot development platform to Intel and SAP's autonomous warehouse practice, it can be seen that physical AI is forming a new infrastructure system. This system requires not only models and computing power, but also sensors, edge computing, communication networks, industrial software, and device management platforms.

For the traditional IoT industry, this means new value space: sensors are no longer just data collection tools, edge computing is no longer just a data preprocessing node, and industrial networks are no longer just device connection channels; together they will support AI's perception, understanding, and autonomous execution of the physical environment.

References:

Google launches its most powerful world model Genie 4, with simulation duration doubled to 2 minutes and audio-video synchronized generation achieved for the first time, gaming stocks face another AI shock—Wallstreetcn

NVIDIA advances the Isaac platform for general-purpose robots with a new humanoid robot reference platform, framework, simulation workflow, and data—NVIDIA

From Task to Fulfillment – Intel and SAP Demonstrate Physical AI and the Autonomous Warehouse in Action—Intel

This article is from the WeChat public account "IoT Think Tank" (ID: iot101), author: Sophia, published by 36Kr with authorization.