Source-linked AI summary
From AI-assisted to AI-native Industrial Control: Architecture, Agent Design, and Challenges
Lisi Liu, Bin Wang, Quan Tang, Chuan Pang, Feng Gao, Shengkai Liu
TL;DR
Traditional AI-assisted ICS adds AI without fundamentally changing the legacy architecture, leaving the potential for greater autonomy and adaptability underused. The paper proposes an AI-native design built around infrastructure and collaborative agents, and demonstrates workflows completed without manual intervention after task prompting. It also identifies reliability, verification, and infrastructure challenges.
Problem
AI-assisted ICS retains traditional architecture, limiting its ability to realize AI’s potential for autonomy, adaptability, and continuous evolution.
Source: §I · p. 1, §II · p. 2Method
The paper proposes an AI-native ICS combining AI-native infrastructure with collaborative agents, including an ICS agent that interprets automation tasks.
Source: A. Architecture of the AI-native ICS · p. 2, B. Key Skills of the ICS Agent and the… · p. 3Results
In both demonstrated cases, all operation phases were completed successfully without manual intervention after the task prompt was submitted.
Source: C. Demonstration of the AI-native ICS · p. 5Takeaways & Limitations
The proposed agent-based architecture and on-demand control-application generation shift ICS toward open, task-oriented solutions synthesized from high-level automation intent.
Source: §V · p. 6AI model hallucinations and limited context reasoning can undermine decision reliability, while AI-generated control applications still require correctness and real-time assurance.
Source: A. Challenges From the AI Perspective · p. 5, B. Challenges From the ICS Perspective · p. 5
Abstract
from arXiv · showhide
The integration of AI and industrial control is rapidly transforming industrial automation. However, current approaches remain predominantly AI-assisted, where AI functions are incorporated as auxiliary components without fundamentally altering the underlying architecture of industrial control systems. Consequently, the potential of AI to enable higher levels of autonomy, adaptability, and continuous evolution remains underutilized. This paper proposes an AI-native industrial control paradigm, which reconceptualizes industrial control systems by positioning AI as the intrinsic core rather than an external enhancement. The proposed paradigm is built upon an AI-native infrastructure and a collection of collaborative agentic AIs, each encapsulating a set of advanced skills. Among them, the industrial control agent plays a central role in orchestrating the system lifecycle. Through this paradigm, industrial control shifts from a closed, capability-oriented, human-driven, and passive engineering process to an open, task-oriented, intent-driven, and autonomous solution delivery mechanism. The paper further analyzes the key challenges of the AI-native paradigm from both AI and industrial control perspectives, including issues related to AI model reliability, formal verification, real-time assurance, and heterogeneous resource orchestration. Overall, this work outlines a forward-looking vision for industrial control in the AI era and identifies critical research directions toward autonomous, evolvable, and adaptive control systems in complex and dynamic industrial environments.
I. INTRODUCTION
AI is entering industrial control through embedded modules and AI-assisted engineering, but these additions leave the traditional system architecture largely intact. The paper argues for redesigning ICS around AI and control as intrinsic capabilities.
I. INTRODUCTION
- Current integration includes AI-in-the-loop modules coupling probabilistic inference with deterministic execution,, and generative-AI copilots for control-program generation and operational guidance –.The paper identifies these as two modalities of AI integration in industrial control.Source: §I · p. 1
- AI-assisted ICS adds AI to legacy systems without redesigning their architecture, limiting autonomy, evolvability, and adaptability.AI remains a peripheral enhancement rather than the system core.Source: §I · p. 1
- The paper argues that mass customization and volatile supply chains require control that executes directives deterministically while responding to emergent physical events,,.It calls for AI and control to be co-designed from first principles rather than retrofitted.Source: §I · p. 1
- The paper proposes an AI-native ICS paradigm and previews its architecture, key components, demonstration, and research challenges.The paper reviews traditional and AI-assisted ICS before introducing the paradigm and discussing challenges from AI and ICS perspectives.Source: §I · pp. 1–2
II. A BRIEF REVIEW OF TRADITIONAL ICS AND AI-ASSISTED ICS
Traditional ICS is organized around closed, vendor-specific PLC systems and a human-driven engineering lifecycle. AI-assisted additions leave this foundation in place, so users still engineer task-specific control programs and the system inherits the traditional limitations.
II. A BRIEF REVIEW OF TRADITIONAL ICS AND AI-ASSISTED ICS
- A PLC couples a vendor-specific IDE for programming, configuration, and diagnostics with an RTE that executes programs on a real-time substrate.Commercial RTEs are typically vertically integrated hardware-software systems designed for determinism, reliability, and safety.Source: §II · p. 2
- Most deployed PLCs are based on IEC 61131-3, which adopts a synchronous scan-cycle execution model, whereas IEC 61499 introduces event-driven execution for reconfigurable and distributed systems.Source: §II · p. 2
- Traditional ICS combines closed, fixed-functionality architecture with passive engineering optimized for stability, contributing to vendor lock-in, poor interoperability, and limited third-party integration.Its vertically integrated ecosystem is sustained by a predefined engineering lifecycle rather than autonomous adaptation.Source: §II · p. 2
- Traditional PLCs provide capability-oriented execution substrates, leaving end users to develop and commission task-specific control programs rather than receiving task-oriented solutions directly.Customized control logic must also be engineered and validated.Source: §II · p. 2
- AI-assisted ICS mainly adds decision support to engineering tools or vendor-specific control interfaces, retaining traditional architecture and its limits on autonomy, adaptability, and continuous evolution.AI augments the legacy system rather than becoming its core.Source: §II · p. 2
III. OVERVIEW OF THE AI-NATIVE ICS PARADIGM
The paper presents AI-native ICS as a first-principles reconceptualization of industrial automation, making AI intrinsic to the system rather than an added feature.
III. OVERVIEW OF THE AI-NATIVE ICS PARADIGM
- AI-native ICS is an intrinsically AI-centered, autonomous, evolutionary, and adaptive reconceptualization of industrial automation.It is framed as a first-principles alternative to adding AI features to existing systems.Source: §III · p. 2
A. Architecture of the AI-native ICS
The AI-native ICS architecture combines AI-native infrastructure with collaborative agentic AIs. Its infrastructure supplies heterogeneous resources and execution support for agents, whose collaboration is associated with autonomy, evolvability, and adaptability.
A. Architecture of the AI-native ICS
- The conceptual architecture pairs AI-native infrastructure with collaborative agentic AIs, treating models and agents as primary executive entities and nondeterminism as inherent and manageable.The infrastructure is organized into physical resources, support functions, and models, runtimes, and tools.Source: A. Architecture of the AI-native ICS · pp. 2–3
- The infrastructure’s physical layer supplies heterogeneous computing, storage, and networking, while its support layer provides inter-agent communication and resource scheduling.A models, runtimes, and tools layer supports specialized agents and their dynamic collaboration.Source: A. Architecture of the AI-native ICS · p. 3
- Collaborating agents such as the ICS, Vision, and Digital Twin agents embody autonomy, anagenesis, and actionability, supporting ICS autonomy, evolvability, and adaptability.The models, runtimes, and tools layer includes AI models, agent runtimes, and utilities for fine-tuning and training.Source: A. Architecture of the AI-native ICS · p. 3
B. Key Skills of the ICS Agent and the Operating Model of the AI-native ICS
The ICS agent orchestrates the AI-native ICS lifecycle through advanced skills, from interpreting automation requirements and deploying control applications to monitoring and continual evolution.
Skill A. Automation task requirements analysis
- It interprets requirements from text, engineering drawings, or formal specifications and transforms them into a unified, machine-executable representation.This representation underpins subsequent agent skills.Source: B. Key Skills of the ICS Agent and the… · p. 3
Skill B. On-demand control application generation and deployment
- The agent turns automation requirements into task-specific control applications, modifying or regenerating them as requirements change and scheduling resources for execution.This shifts from selecting a predefined RTE and coding for it to a task-oriented solution.Source: B. Key Skills of the ICS Agent and the… · p. 3
Skill C. Legacy system compatibility
- For backward compatibility, the agent can select a legacy RTE configuration, generate installation guidance, and produce a tailored control program.This allows the AI-native paradigm to subsume AI-assisted functions when needed and offers a non-disruptive migration path.Source: B. Key Skills of the ICS Agent and the… · p. 3
Skill D. Agents interoperation
- The ICS agent can coordinate with specialized agents, including industrial robot agents that perform installation and wiring, and respond to requests from other agents.This supports cooperative automation across the industrial landscape.Source: B. Key Skills of the ICS Agent and the… · p. 4
Skill E. Virtual commissioning
- Before deployment, the agent uses a digital twin agent to validate, debug, and optimize control logic in a virtual environment.The paper says virtual commissioning reduces on-site commissioning time, cost, and risks.Source: B. Key Skills of the ICS Agent and the… · p. 4
Skill F. Proactive monitoring and maintenance
- After deployment, the agent monitors the live system, detects anomalies, predicts failures, and can trigger maintenance procedures.Maintenance can be coordinated with other agents.Source: B. Key Skills of the ICS Agent and the… · p. 4
Skill G. Anagenesis: continual skill and performance evolution
- Anagenesis lets the ICS agent learn from operational data, simulations, and agent interactions to refine existing skills and develop new ones.The paper contrasts this evolution with traditional ICS’s fixed-function nature.Source: B. Key Skills of the ICS Agent and the… · p. 4
C. Demonstration of the AI-native ICS
The demonstration reports that both sorting cases completed all operation phases without manual intervention after a task prompt. It illustrates AI as a control architect while deterministic logic governs actuators.
Sorting scenario
- The demonstration covers single-product sorting to a predefined location and multi-product sorting using camera-based identification and corresponding destinations.The cases are illustrated with finite-state-machine control logic.Source: C. Demonstration of the AI-native ICS · p. 4, Figure 5 · p. 4
Demonstration results
- Both sorting cases completed all operation phases without manual intervention after the task prompt was submitted.The paper compares the AI-native workflow with conventional PLC-based engineering.Source: C. Demonstration of the AI-native ICS · p. 5
Control architecture
- The ICS agent acts as an autonomous control architect that translates tasks into deterministic, verifiable applications, while actuators remain governed by deterministic logic.The paper distinguishes this from assigning AI direct real-time control.Source: C. Demonstration of the AI-native ICS · p. 5
IV. CHALLENGES OF AI-NATIVE ICS PARADIGM
The AI-native paradigm introduces challenges through deep AI–ICS integration, which the paper examines from both AI and industrial-control perspectives.
Challenges of AI-native ICS
- Deep integration of AI and ICS introduces a new set of challenges for the proposed paradigm.The passage does not specify individual challenges.Source: §IV · p. 5
AI perspective
- The paper examines the challenges from the AI perspective.Source: §IV · p. 5
ICS perspective
- The paper also examines the challenges from the industrial-control perspective.Source: §IV · p. 5
A. Challenges From the AI Perspective
AI-native ICS depends on advanced models and agentic runtimes, but their reliability and the immaturity of supporting infrastructure remain challenges.
A. Challenges From the AI Perspective
- Hallucination and limited context reasoning in advanced AI models may make decisions unreliable on long-horizon or complex tasks, with risks extending to control failures or physical accidents.The paper identifies verifiability, predictability, and bounded behavior as basic requirements for agentic AI in ICS.Source: A. Challenges From the AI Perspective · p. 5
- AI-native infrastructure remains at an early stage without unified architectures or standards, and its layered architecture is only a conceptual reference.Open issues include agent standardization, communication mechanisms, training and inference toolchains, and resource scheduling across heterogeneous infrastructure.Source: A. Challenges From the AI Perspective · p. 5
B. Challenges From the ICS Perspective
AI-native ICS envisions control logic generated and deployed as requirements change, but traditional closed systems are built for stability and tightly controlled execution. This creates challenges in verifying generated applications and deploying them across heterogeneous platforms.
B. Challenges From the ICS Perspective
- Continuous generation and deployment of control logic conflicts with traditional ICS built around closed, vendor-specific environments, stability, and predefined functionality.This mismatch motivates the ICS agent’s on-demand control-application generation and deployment skill.Source: B. Challenges From the ICS Perspective · p. 5
- AI-generated control applications must meet strict determinism, latency, execution-time, and functional-correctness requirements.The paper identifies formal modeling and verification, including timed or hybrid automata, as promising ways to check timing constraints and reachability.Source: B. Challenges From the ICS Perspective · p. 5
- Control applications may be decomposed, distributed, or migrated across platforms, but hardware coupling and strict real-time constraints make dynamic reallocation difficult.Virtualization can decouple control logic from hardware, yet CPU architectures, memory hierarchies, and accelerators remain practical constraints.Source: B. Challenges From the ICS Perspective · p. 5
C. Summary
AI-native ICS faces a basic tension between evolving probabilistic AI and deterministic, safety-critical control. Addressing it requires verification, real-time resource governance, and standardized infrastructure.
C. Summary
- The paradigm’s core challenge is reconciling probabilistic, evolving AI with the deterministic, safety-critical requirements of ICS.AI reliability concerns include hallucination, limited context reasoning, and the absence of verifiable guarantees; the ICS transition also raises control-logic verification and resource-orchestration challenges.Source: C. Summary · p. 5
- The paper identifies formal verification, real-time resource governance, and standardized AI-native infrastructure as needed foundations for autonomous, evolvable, and adaptive ICS.These challenges reflect the deep integration of AI and ICS rather than an incremental extension of traditional systems.Source: C. Summary · p. 5
V. CONCLUSION
The conclusion frames AI-native ICS as a first-principles redesign that makes AI intrinsic to control and uses an agent-based architecture supported by AI-native infrastructure. It also highlights verification, real-time reliability, and resource orchestration challenges.
V. CONCLUSION
- The proposed paradigm treats AI as the intrinsic core of ICS, rather than an external enhancement to legacy architecture.Table I is identified as an experimental validation comparison between conventional PLC and proposed AI-native ICS.Source: §V · p. 6, Table I · p. 6
- An agent-based architecture supported by AI-native infrastructure is proposed, with the ICS agent generating and deploying task-specific applications from high-level automation intent.This shifts the system from a closed, capability-oriented solution toward an open, task-oriented one.Source: §V · p. 6
- The paper identifies verifiable and bounded AI behavior, formalized control logic, real-time determinism, and heterogeneous-resource orchestration as research challenges.It calls for AI-native infrastructure and trustworthy verification and resource-orchestration mechanisms.Source: §V · p. 6