12 Software Layers Required for Physical AI

Humanoid Development Roadmap: Software Edition, Part 2

A Comprehensive Map for Transforming Humanoids from "Walking Machines" into "Working AI"

In the previous article, we outlined why software, not just hardware, is critical in physical AI development.

The theme of the first installment was a sense of urgency and strategy.

Japan is globally strong in motors, reducers, sensors, industrial robots, machine tools, and precision components. However, in an era where humanoids are fully implemented in society, the center of value will shift from hardware to software.

A robot that can walk is different from a robot that can work.

Robot bodies will become purchasable. We are entering an era where general-purpose robot data can also be bought. However, failure data, safe operation data, and operational integration data that can only be obtained at your own company's site must be cultivated in-house.

In the first installment, we organized these major premises.

In this second installment, we will take a step further.

This time, we will organize the software required to actually develop and operate physical AI by dividing it into 12 layers.

In other words, if the first installment was an article about the philosophy and strategy of "why software is important," this second installment is a comprehensive map showing what software layers are actually needed.

The software that powers a humanoid is not a single AI model.

Simply inserting a large language model like ChatGPT does not mean a robot can work on its own.

Simply inserting a VLA model does not mean it can perform housework or factory tasks on its own.

Simply using a world model does not mean it can perfectly understand the real physical world on its own.

Physical AI requires multiple software layers.

Only when these are stacked together does a humanoid transform from a "walking machine" into a "working physical AI."

1. Why is it necessary to divide it into 12 layers?

When talking about physical AI, many people focus on one prominent technology.

For example, VLA. For example, world models. For example, simulation.

Of course, all of these are important.

However, physical AI cannot be established by a single technology alone.

Robots are AI that operate in the real world.

The real world has friction. It has gravity. It has contact. It has collisions. It has lighting changes. It has sensor errors. It has unpredictable human movements. It has individual equipment differences. It has site-specific rules. It has safety standards. It has business systems.

Therefore, physical AI requires a multi-layered software structure.

It is easy to understand if you compare it to humans.

Humans have eyes. They have ears. They have skin. They have a sense of balance. They have a cerebrum. They have a cerebellum. They have spinal reflexes. They have memory. They have experience. They have social rules. They have workplace rules.

Robots also need multiple layers in the same way.

It is not realistic for a single massive AI model to safely handle all of these.

That is why it is important to understand physical AI by dividing it into 12 layers.

2. 12 Software Layers Required for Physical AI

The 12 layers organized here are as follows.

Number Software Layer Role 1 Perception Layer Reads the real world from sensors 2 World Model Layer Predicts the structure of the world and the future 3 Cerebrum AI Layer Performs language understanding, reasoning, planning, and task decomposition 4 VLA Layer Sees, understands language, and outputs actions 5 Skill Library Layer Makes basic movements into reusable components 6 Cerebellum Control Layer Controls posture, walking, grasping, and force application 7 Robot OS/Middleware Layer Connects sensors, AI, control, and actuators 8 Digital Twin Layer Recreates the actual robot and site in a virtual space 9 Simulation/Sim-to-Real Layer Trains in a virtual space and transfers to reality 10 Data Infrastructure Layer Accumulates actual machine logs, failure data, and learning data 11 Safety/Security Layer Protects people, equipment, and robots 12 RoboOps/Site Coordination Layer Performs multi-unit operation, remote monitoring, and ERP/MES/WMS integration

These 12 layers do not simply flow in one direction from top to bottom.

In reality, they are interrelated.

In other words, physical AI is not a single AI model, but a circulating software system.

See. Understand. Think. Move. Fail. Record. Re-learn. Operate safely.

Creating this cycle is the essence of physical AI development.

Note that the groups of companies listed for each layer are organized as "major players" or "benchmark companies." This is not intended to define the world's top companies, but rather to be used as a reference for representative examples that are making noteworthy moves.

3. Layer 1: Perception Layer - Reading the real world from sensors

The first layer is the perception layer.

For a robot to move in the real world, it must first be able to read the world.

Cameras alone are not enough. LiDAR alone is not enough. IMU alone is not enough. Tactile sensors alone are not enough.

A robot operating in the real world requires a perception system that combines multiple sensors.

Data entering the perception layer

• RGB camera images

• Depth camera images

• Stereo cameras

• LiDAR point clouds

• IMU data

• Encoder information

• Force sensors

• Tactile sensors

• Foot pressure sensors

• Microphone audio

• Distance sensors

Processing performed in the perception layer

Object detection, image classification, segmentation, depth estimation, pose estimation, human detection, 3D object recognition, SLAM, localization, sensor fusion, contact detection, slip detection, and anomaly detection.

In short, the perception layer is the layer that converts "raw sensor data" into "meaningful state information."

If this is weak, neither the subsequent world model nor the VLA will function correctly.

If the input is wrong, even the smartest cerebral AI will make mistakes.

The first step for Physical AI is to read the real world correctly.

Key Player Map

• Sensor Components: Keyence (Japan), Omron (Japan), Sony (Japan, CMOS image sensors), Hokuyo (Japan, LiDAR), Ouster (post-Velodyne integration), Livox (LiDAR), RealSense (formerly Intel RealSense, spun out from Intel as an independent company in July 2025, depth cameras)

• Tactile Sensors: GelSight (USA), Sanctuary AI (Canada), ifm electronic (Germany)

• 3D Vision/SLAM: Boston Dynamics, Skydio, Mobileye, Wayve

Target Companies for Japanese Firms: This is a layer where Japanese companies can easily leverage their strengths in components, sensors, and image processing. Keyence, Omron, and Sony, in particular, have a global presence. On the other hand, overseas ecosystems are ahead in the software layer that converts sensor data into state information for robots.

4. Layer 2: World Model Layer—Predicting the Structure and Future of the World

If the cognitive layer is the layer that reads the real world, the world model layer is the layer that understands that world and predicts the future.

The world model is the core of Physical AI.

However, a world model is not just image recognition.

It is necessary not only to recognize what is in front of you, but also to predict how that object will move, whether it will fall if pushed, and where a person will move next.

Difference between World Models and Physics Simulators

A physics simulator is an 'environment that calculates physical laws.' A world model is an 'AI that has learned how the world changes.'

These two are not contradictory. They are used in combination.

Key Player Map

• World Foundation Model: NVIDIA (Cosmos Predict / Transfer / Reason), Google DeepMind (Genie 3; Veo 3 is a related video generation model), World Labs (Marble, founded by Fei-Fei Li)

• Physics Simulator-based World Models: Boston Dynamics, Tesla, Wayve (for autonomous driving)

• Interactive World Generation: Decart, Odyssey, Runway

Target Companies for Japanese Firms: This is a layer where a head-on confrontation is difficult. It is realistic to build world models specialized for the field while using foundations from NVIDIA or World Labs. Details will be covered in the 3rd and 4th installments.

5. Layer 3: Cerebral AI Layer—Language Understanding, Reasoning, Planning, and Task Decomposition

Understanding the world alone is not enough for a robot to work.

The next requirement is a layer that thinks about what needs to be done.

This is the Cerebral AI Layer.

It understands instructions from humans, interprets goals, breaks them down into work procedures, selects necessary skills, and determines the execution order.

For a home robot, an instruction like "tidy up the living room" must be broken down into steps: pick up items on the floor, sort trash, return books to the shelf, carry dishes to the kitchen, avoid touching dangerous items, and perform a final check.

This is a higher-level process than VLA.

Key Player Map

• Large Language Models: OpenAI, Anthropic, Google DeepMind, Meta, xAI

• Chinese Firms: DeepSeek, Qwen (Alibaba), Doubao (ByteDance), Baidu ERNIE

• Integration into Robots: Figure AI (transitioned to Helix after utilizing GPT-based systems), 1X Technologies, Skild AI

Target Companies for Japanese Firms: It is difficult to compete head-on in the foundation model race here as well. In Japan, companies like Preferred Networks, ELYZA, Sakana AI, NTT (tsuzumi), and Fujitsu (Takane) are advancing foundation model development. The target should not be to use general-purpose LLMs as they are, but to create Cerebral AI for robots fine-tuned with Japanese on-site manuals, safety rules, and operational knowledge.

6. Layer 4: VLA Layer—Seeing, Understanding Language, and Generating Actions

The VLA layer is an area receiving significant attention within Physical AI.

VLA stands for Vision-Language-Action.

When given an instruction like "pick up the blue cup," the VLA layer finds the blue cup in the camera image and generates an action directed at that object.

For VLA, there are two approaches: an end-to-end approach that outputs directly to joint movements, and a hierarchical approach where the VLA only issues high-level action instructions, while the skill library and cerebellar control underneath handle the actual movements.

The industry has not yet reached a consensus on which is the correct approach.

Key Player Map

• VLA Specialized Startups: Physical Intelligence (π0 series, with π0.7 also announced as of 2026, led by Alphabet's CapitalG), Skild AI (Skild Brain)

• In-house VLA of humanoid manufacturers: Figure AI (Helix, System 1/System 2 configuration; System 0 will also be added in the 2026 Helix 02), Tesla (in-house model for Optimus), 1X Technologies, AgiBot (GO-1)

• Foundation model providers: NVIDIA (GR00T N1.5, N1.7), Google DeepMind (Gemini Robotics 1.5)

• Open source: OpenVLA (Stanford), π0 / π0.5 (partially open-sourced by Physical Intelligence)

Target companies for Japanese firms: Aiming to create Japan-specific, field-specialized VLAs while benchmarking against Physical Intelligence and Skild AI. Preferred Networks and Sakana AI are leading Japanese candidates, but they are still in the development stage.

7. Layer 5: Skill Library Layer—Making basic movements into reusable components

Even if a VLA or cerebral AI makes high-level decisions, a robot needs basic movements it can execute.

This is the skill library layer.

Grasping, placing, pushing, pulling, turning, opening, closing, carrying, inserting, avoiding, handing over, receiving, wiping, folding.

These are basic skills turned into components.

Without a skill library, a VLA or LLM would have to think through low-level movements every single time.

This is extremely unstable.

In practical physical AI, it is important to design the system so that the higher-level AI calls verified skills rather than controlling everything directly.

Major Player Map

• General-purpose skill library: NVIDIA Isaac Manipulator, Physical Intelligence (π0 series), Skild AI (Skild Brain)

• Industry-specific skills: ABB Robotics, FANUC (Japan), Yaskawa Electric (Japan), KUKA, Universal Robots, Kawasaki Heavy Industries (Japan)

• Manipulation specialists: Covariant (Amazon acquired a non-exclusive license for their robot foundation model in 2024 and hired the founders and some staff), Sanctuary AI, Dexterity

Target companies for Japanese firms: This is one of the layers where Japanese companies can easily compete. FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Denso, and Mitsubishi Electric have accumulated basic industrial robot skills over many years. If they can convert these into skill libraries for the humanoid era, they have the potential to compete globally.

8. Layer 6: Cerebellum Control Layer—Controlling posture, walking, grasping, and force application

The cerebellum control layer is the layer for moving the robot's body stably.

Arm trajectory, joint angles, torque, center of gravity, foot contact, end-effector force, and posture stability.

The cerebellar control layer handles these.

Posture control, gait control, balance control, inverse kinematics, forward kinematics, dynamics, motion planning, MPC, whole-body coordination control, impedance control, force control, visual servoing, reflex control, fall avoidance, grasping control, and finger control.

In humanoids, whole-body coordination is particularly difficult.

No matter how smart the cerebral AI is, it cannot move in the real world if the cerebellar control is weak.

Key Player Map

• Representative players in gait and whole-body control: Boston Dynamics (Atlas, under Hyundai), Unitree (H2), Agility Robotics (Digit)

• Finger control: Sanctuary AI (Phoenix), Shadow Robot (UK), Clone Robotics

• MPC and force control: Various spin-outs from ETH Zurich, Boston Dynamics

• Industrial control: FANUC (Japan), Yaskawa Electric (Japan), ABB, KUKA, Universal Robots

Target companies for Japanese firms: This area is a field where Japanese robot control technology is traditionally strong. FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Denso Wave, and Mitsubishi Electric have global track records in industrial robots and control technology. If they advance whole-body coordination control for humanoids, they could potentially compete with Boston Dynamics and Unitree. Honda Motor Co., Ltd. (the accumulation of ASIMO development), the National Institute of Advanced Industrial Science and Technology (AIST), the University of Tokyo (JSK Lab), and AIST's HRP series are also intellectual property of Japanese cerebellar control.

9. Layer 7: Robot OS/Middleware Layer—Connecting sensors, AI, control, and actuators

A representative example is ROS 2.

In physical AI, sensors, AI models, control software, motors, logs, communication, visualization, and simulation are complexly connected.

If you run these separately and in a fragmented way, they will quickly fail.

Therefore, middleware that connects the entire robot system is necessary.

To use a human analogy, it is the "wiring" within the nervous system.

However, ROS 2 is not the core of real-time control itself. It is common for low-layer humanoid control (joint control loops of 1kHz or higher) to run on a separate real-time control layer, not on top of ROS 2.

Key Player Map

• Core of ROS / ROS 2: OSRF (Open Source Robotics Foundation, the non-profit management body for ROS/Gazebo, etc.), Intrinsic (under Alphabet, acquired former OSRC/OSRC-SG assets)

• DDS implementation: Eclipse Cyclone DDS, eProsima Fast DDS, RTI Connext

• micro-ROS / Real-time control: eProsima, ROS-Industrial Consortium

• Commercial Robot OS: NVIDIA Isaac ROS, Intrinsic Flowstate, Apex.AI (for automotive)

Target companies for Japanese firms: This is an overseas-led domain. In Japan, while there are contributions from the ROS-Industrial Consortium Japan Chapter, TIER IV (Autoware for autonomous driving), and various universities (University of Tokyo, Osaka University, AIST), the global foundation is not held by Japan. Japanese companies should aim to master ROS 2 while creating middleware extensions for specific field applications on top of it.

10. Layer 8: Digital Twin Layer—Recreating real robots and sites in virtual space

The digital twin layer is a layer that recreates real robots, factories, warehouses, homes, equipment, and human movement paths in virtual space.

What is important here is that a digital twin is not just 3D CG.

What is needed is a virtual environment where robots can actually move.

To achieve this, physical attributes such as friction, mass, collision shapes, range of motion, coefficient of restitution, distinction between rigid and soft bodies, sensor noise, and lighting conditions are also necessary.

For buildings and factories, BIM is the starting point. For equipment and machinery, CAD is important. For existing facilities, 3D scanning and point cloud data are required.

Major Player Map

• Digital twin infrastructure for robots: NVIDIA Omniverse / Isaac Sim (OpenUSD-based, one of the most prominent foundations)

• Industrial digital twin: Siemens (Plant Simulation, Tecnomatix), Dassault Systèmes (3DEXPERIENCE), PTC (Onshape to Isaac Sim integration), Hexagon, Bentley Systems

• 3D spatial generation AI: World Labs (Marble), Niantic Spatial, Autodesk (in collaboration with NVIDIA)

• BIM: Autodesk (Revit), Bentley, Trimble

Target companies for Japanese firms:This layer offers significant opportunities for Japanese companies. Japanese companies that possess BIM, CAD, and 3D scan data for factories, warehouses, and facilities are in a position to convert them into digital twins for robots. Hitachi, Fujitsu, NEC, Sharp, Daikin, FANUC, Yaskawa, Denso, Shimizu Corporation, Obayashi Corporation, and Takenaka Corporation all possess BIM/CAD assets. It is realistic to utilize these in combination with NVIDIA Omniverse or World Labs Marble.

11. Layer 9: Simulation/Sim-to-Real Layer—Training in virtual space and transferring to reality

This is the layer where robots are operated, trained, and verified within a digital twin.

In the real world, you cannot have a robot fall over tens of thousands of times.

However, you can do this in simulation. In actual humanoid reinforcement learning, hundreds of thousands to millions of episodes are run in virtual space.

Sim-to-Real refers to transferring skills or policies learned in simulation to actual real-world hardware.

However, there is a discrepancy between simulation and reality. This is called the Sim-to-Real Gap.

To bridge this gap, Domain Randomization, Real-to-Sim, System Identification, sensor noise models, real-machine log collection, failure data reproduction, and retraining are necessary.

Key Player Map

• Simulation Infrastructure: NVIDIA Isaac Sim / Isaac Lab, Google DeepMind (MuJoCo, open source), Coppelia Robotics (CoppeliaSim), Open Robotics (Gazebo)

• Sim-to-Real Specialized Research: OpenAI (past research), ETH Zurich, UC Berkeley (BAIR), CMU, Stanford, University of Tokyo JSK Lab

• Synthetic Data/Domain Transfer: NVIDIA (Cosmos Transfer), Lightwheel (for Samsung), PTC (Onshape to Isaac Sim)

Target Companies for Japanese Firms: Since Isaac Sim / Isaac Lab is one of the most promising foundations, it is realistic to focus on mastering it. As for unique Japanese simulation foundations, there are AIST's Choreonoid and simulation research related to the University of Tokyo.

12. Layer 10: Data Infrastructure Layer—Accumulating real-machine logs, failure data, and training data

Robots generate data every time they move.

If you discard this data, the robot will not become smarter.

Conversely, if you have a system to accumulate, analyze, and use this data for retraining, the robot will improve the more it is used.

General-purpose data and company-specific data are different

As organized in the first installment, there is a possibility that general-purpose robot motion data and synthetic data will be available for purchase from external sources in the future.

Grasping data, walking data, teleoperation data, shelf work data, and synthetic data generated by simulation.

However, that alone is not enough to enter the field.

Your company's equipment, your company's workers, your company's shelf layout, your company's flooring, your company's lighting, your company's safety rules, your company's work standards, and your company's failure patterns.

Data tied to these must be accumulated by operating robots at your own site.

Build initial capabilities with general-purpose data. Adapt to the site with company-specific data. Improve with failure data. Reduce risk with safe operation data.

Companies that can create this flow will be strong in the era of Physical AI.

Key Player Map

• Cloud MLOps Infrastructure: AWS SageMaker, Azure ML, Google Vertex AI, Databricks

• Robot Operations Data Infrastructure: Agility Arc (Agility Robotics), Formant, Foxglove (open-source ROS visualization), InOrbit

• Public Robot Datasets: Open X-Embodiment (Google-related), DROID, RT-2 Dataset, AgiBot Dataset, NVIDIA Open Physical AI Dataset

• Data Generation Infrastructure: NVIDIA Physical AI Data Factory Blueprint, Skild AI, Lightwheel

Target Companies for Japanese Firms: Data infrastructure is an area where Japanese companies can easily combine traceability, quality control, and a culture of improvement. Fujitsu, NEC, Hitachi, NTT DATA, and ITOCHU Techno-Solutions have experience in cloud and data infrastructure. By combining this with on-site data from Japanese robot manufacturers (FANUC, Yaskawa), a strong position can be secured.

13. Layer 11: Safety and Security Layer—Protecting People, Equipment, and Robots

In Physical AI, safety design is essential.

This is because robots operate in the real world.

The type of risk involved when generative AI makes a mistake in text is different from when a humanoid bumps into a person.

Emergency stop, human detection, collision avoidance, speed limits, force limits, torque limits, hazardous area recognition, workspace restrictions, fall detection, abnormal posture detection, AI behavior filters, Safety Shield, log recording, and accident reconstruction.

The safety layer should be the final barrier placed on top of the AI model.

Safety is not something to be added at the end

What is important here is that the safety and security layer is not something to be retrofitted at the end of development.

Safety philosophy must be incorporated into all layers from the very beginning.

In other words, safety is both an independent function and a cross-cutting concern that applies to all layers.

Key Player Map

• Functional Safety Certification: TÜV SÜD (Germany), TÜV Rheinland (Germany), UL Solutions (USA), SGS (Switzerland), JQA (Japan)

• Industrial Robot Safety Standards: ISO 10218-1/-2:2025, ISO/TS 15066 (related specifications for collaborative robot safety; some elements integrated into the 2025 version of ISO 10218), IEC 61508, ISO 13849

• Robot Security: Karamba Security, Forescout, Claroty (OT security)

• AI Safety: Safety Teams at Anthropic, OpenAI, and DeepMind

Target Companies for Japanese Firms:This is also an area with strong opportunities for Japanese companies. Japan has a wealth of accumulated expertise in functional safety, ISO certification, and on-site safety operations. Certification and standardization bodies such as the Japan Electrical Control Equipment Industries Association (NECA), the Japan Robot Association (JARA), AIST, JET, and JQA. By combining this with know-how in on-site safety operations (KYT, near-miss reporting, and risk assessment), there is potential to lead the world in the safety layer of physical AI.

14. Layer 12: RoboOps and On-site Integration Layer—Multi-unit operation, remote monitoring, and ERP/MES/WMS integration

The final layer is the RoboOps and On-site Integration Layer.

If there is only one unit, a human can watch it from the side.

However, when it comes to 10, 100, or 1,000 units, an operational foundation becomes necessary.

Remote monitoring, fleet management, utilization management, anomaly detection, log collection, model updates, OTA updates, predictive maintenance, battery management, work history management, safety reporting, maintenance management, and human operator intervention.

RoboOps is the robot version of DevOps and MLOps.

On-site System Integration

MES for factories. WMS for warehouses. ERP for the entire enterprise. QMS for quality control. EAM for facility maintenance.

For a robot to know 'what it should do,' it must receive instructions from on-site systems.

Key Player Map

• Robot Fleet Management (RaaS Infrastructure): Agility Arc (Agility Robotics), Formant, Freedom Robotics (Roboto), InOrbit, Cobalt AI

• MES: Siemens Opcenter, Rockwell Automation, Dassault DELMIA, SAP Manufacturing

• WMS: Manhattan Associates, Blue Yonder, Körber, SAP EWM, Oracle WMS

• ERP: SAP, Oracle, Microsoft Dynamics

• Japanese-affiliated: Fujitsu (GLOVIA), Hitachi (HITPHAMS), NEC (EXPLANNER), Toshiba (MES Bridge)

Target Companies for Japanese Firms:This layer is one of the areas that Japanese companies should target the most. Japanese manufacturing and logistics sites have extensive experience in implementing proprietary MES, WMS, ERP, and QMS. System integrators such as Fujitsu, NEC, Hitachi, NTT DATA, ITOCHU Techno-Solutions, SCSK, TIS, and Sumitomo Electric Information Systems can play the role of connecting on-site systems with robots.

15. Comparing the 12 layers to the human body

Human Physical AI Software Layers Eyes/Ears/Skin Cognitive Layer World Understanding in the Head World Model Layer Cerebrum Cerebrum AI Layer Ability to turn vision into words and actions VLA Layer Learned Action Skill Library Layer Cerebellum/Reflex Cerebellar Control Layer Nervous System ROS 2/Middleware Layer Practice Field in the Head Digital Twin Layer Iterative Practice Simulation/Sim-to-Real Layer Memory/Experience Data Infrastructure Layer Danger Avoidance/Immunity Safety/Security Layer Connection to Society/Organization RoboOps/On-site Integration Layer

Thinking about it this way, it becomes clear that physical AI is not just a simple AI model.

Robots are systems that possess a physical body.

Therefore, like humans, they require sensation, understanding, judgment, movement, memory, safety, and social connectivity.

16. Development Order of the 12 Layers

Phase 1: Basic Implementation

Robot OS/Middleware layer, Cerebellar control layer, Cognitive layer, Digital twin layer, and Simulation layer.

First, create a state where the robot can see, move, record, and be tested in a virtual space.

Phase 2: Intelligence

World model layer, VLA layer, Cerebrum AI layer, and Skill library layer.

Here, we build the mechanisms for seeing, thinking, acting, and combining basic movements.

Phase 3: Field Implementation

Data infrastructure layer, Safety/Security layer, and RoboOps/Field coordination layer.

However, safety philosophy must be integrated throughout from Phase 1.

Only after achieving this can we move from research and development toward social implementation.

17. Which layers are easily overlooked

There are layers that are easily overlooked in physical AI development.

• Skill library layer: While attention is focused on VLAs and LLMs, the layer that modularizes actual basic movements tends to be neglected.

• Data infrastructure layer: Even if it works in a demo, you cannot improve without accumulating failure data.

• Safety/Security layer: As AI becomes smarter, there is a possibility it may take dangerous actions.

• RoboOps/Field coordination layer: A demo with one unit and field operations with 100 units are completely different things.

18. Layers that Japanese companies can easily target

In the first installment, I outlined why Japan should focus on the software domain of Physical AI.

In this second installment, I will more specifically organize which of the 12 layers are most accessible for Japanese companies to target.

Targetable Layer | Reason | Benchmark | Overseas Companies | Perception Layer (Sensor Fusion) | Keyence, Omron, and Sony are strong in components | Mobileye, Skydio | Skill Library Layer | Accumulation of Fanuc, Yaskawa, and Kawasaki Heavy Industries | Physical Intelligence, Skild AI | Cerebellar Control Layer | Global track record in industrial robot control | Boston Dynamics, Unitree, Agility | Digital Twin Layer | Good compatibility with factories, equipment, BIM, and CAD | NVIDIA Omniverse, Siemens, Dassault | Data Infrastructure Layer | Good compatibility with quality control and traceability | Agility Arc, Formant | Safety & Security Layer | Good compatibility with Japanese safety culture and functional safety | TÜV, UL, Karamba Security | RoboOps & Field Integration Layer | Maintenance, operations, and ERP/MES/WMS/QMS integration | Manhattan Associates, SAP, Siemens

It is not easy for Japanese companies to suddenly create the world's largest VLA or world models.

However, they can demonstrate strengths in layers closer to the field.

The race to mass-produce humanoid bodies at the lowest cost may be led by Chinese companies.

The foundation model race may be led by massive US and Chinese corporations.

However, in the areas of deploying to the field, connecting to operations, operating safely, and continuously cycling improvement data, Japanese companies have significant potential.

19. Summary: Physical AI is a Comprehensive System of 12 Layers

Physical AI is not just a single AI model.

It is not just a world model. It is not just VLA. It is not just LLM. It is not just Isaac Sim. It is not just ROS 2.

It is a comprehensive system of 12 layers that combines them all.

In other words, the essence of Physical AI is raising robots through software.

In the first installment, I outlined why software is important.

In this second installment, I divided that software into 12 layers and mapped it out.

Starting next time, I will delve deeply into the core technologies that are particularly important among these 12 layers, one by one.

Next Preview

Next time, I will cover “What is a World Model: The Technology to Create a ‘Real World in the Mind’ for Robots”.

The world model is the core of Physical AI.

It is the technology that allows a robot not only to recognize objects in front of it, but also to predict how those objects will move, what will happen if it pushes them, whether they will slip if it grabs them, where a person will move next, and whether its own actions will succeed or fail.

Next time, I will explain what the "real world inside the head" means for a robot, while clarifying the differences between world models, physical simulations, digital twins, and VLAs.

*Information on the key players in each layer of this article is organized based on content published and reported as of the first half of 2026. This does not definitively identify the "world's top" players, but should be used as a reference for representative examples that are making notable moves.

#PhysicalAI #Humanoid #PhysicalAI #Robotics #AI #12Layers #SoftwareArchitecture #CognitiveLayer #WorldModel #VLA #SkillLibrary #CerebellarControl #ROS2 #DigitalTwin #SimToReal #DataInfrastructure #SafetyDesign #RoboOps #FANUC #YaskawaElectric #Keyence #Omron #Sony #BostonDynamics #AgilityRobotics #Unitree #NVIDIA #IsaacSim #Omniverse #PhysicalIntelligence #SkildAI #Siemens #Dassault #Anthropic #FieldImplementation #JapaneseManufacturing #HardwareSoftwareIntegration #USChinaConflict #FutureOfJapan