Siemens Adds Guardrail Layer to Keep AI From Overloading Factory PLCs
Ai4
Industrial autonomy requires deterministic state orchestration below SCADA because high-frequency model writes overwhelm legacy controllers while language models fail to understand physical plant causality.
## Background
Published on YouTube by Ai4 on August 23, 2026, this presentation features Bryan Albrecht and Caleb Eastman from Siemens Digital Industries discussing agentic AI operations and data orchestration in manufacturing. The session explores how industrial enterprises can transition from siloed AI pilots to production-scale, closed-loop systems that safely sense, decide, and act in real time while preserving human operator control.
## TL;DR
Bryan Albrecht and Caleb Eastman outline Siemens' architectural blueprint for autonomous industrial operations through Smart Production Orchestration (SPO). They emphasize moving beyond basic LLM harness guardrails to incorporate multi-actor coordination and physics-based operational state machine guardrails that prevent automation from violating physical machine constraints. High-frequency AI micro-adjustments risk overwhelming legacy PLC controllers and networks, requiring deterministic governance layers below SCADA that issue time-bounded control leases. Ultimately, sustainable scaling demands closed-loop feedback systems inspired by autonomous vehicle architectures, treating LLMs not as causal reasoning engines but as structured translation interfaces for human operators.
## Atmosphere & style
The presentation carries a grounded, authoritative, and refreshingly no-nonsense tone that strips away AI hype in favor of hard-nosed operational reality. Albrecht and Eastman display an easy, collaborative chemistry, enlivening complex systems engineering concepts with relatable midwestern agricultural analogies like laser weeding. Listening to the session feels like an insider masterclass in factory-floor pragmatism, balancing high-level architectural ambition with an empathetic respect for plant operators and the physical limits of industrial hardware.
## Key takeaways
- [01:20](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=80s) Industrial autonomy requires a three-tier guardrail architecture: model-level LLM harness guardrails, multi-actor coordination guardrails, and operational guardrails tied to hierarchical physical state machines.
- [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) Operator trust is the ultimate adoption bottleneck; manufacturing autonomy projects fail if frontline operators reject the interface during initial use.
- [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) Smart Production Orchestration uses time-bounded control leases and dynamic envelopes to prevent stale model decisions from executing on PLCs.
- [09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) Industrial operations are shifting from macro-level manual adjustments to high-frequency micro-adjustments, analogous to the evolution from broadcast spraying to targeted laser weeding in agriculture.
- [09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) Failing to account for machine settling times during high-frequency AI setpoint adjustments can saturate PLC processors and overwhelm legacy OT network infrastructure.
- [13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) Operational technology (OT) governance fundamentally differs from corporate IT governance by directly managing physical safety, high-speed equipment, and deterministic machine boundaries.
- [13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) To achieve closed-loop control without latency bottlenecks, the orchestration layer must sit below traditional SCADA/DCS supervisory systems.
- [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) End-to-end industrial autonomy requires three distinct loops: real-time control execution, data-to-training feedback, and continuous operational drift management.
- [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) Long-term operational advantage mirrors autonomous vehicle fleets, where value comes from building an 'AI factory' feedback pipeline that continuously learns from production data.
- [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) Many manufacturing challenges are best solved with deterministic engineering and traditional logic rather than applying machine learning indiscriminately.
- [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) Large language models should not be treated as causal reasoning engines on the factory floor, but rather as translation layers between human operator intent and structured machine telemetry.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=4s) Introduction and the Push Toward Autonomous Manufacturing
Bryan Albrecht, Director of deployment teams for Smart Manufacturing Digital Thread & Incubation and Validation Deployment (SMDT & IVD) at Siemens Digital Industries, introduces the session alongside Caleb Eastman, Field CTO for the Americas at Siemens. Eastman explains his role plainly: whenever a customer across the Americas wants to implement autonomy on their manufacturing floor, his and Albrecht's deployment teams get the call. Before diving into the technical mechanics, Albrecht takes a quick poll of the audience to gauge how many attendees are currently actively pursuing fully autonomous manufacturing systems, noting a substantial show of hands across the room.
[01:20](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=80s) The Three Tiers of Industrial Guardrails: Harness, Coordination, and Operational
While data orchestration has long been requested by industrial enterprises, modern agentic systems require deep operational context and strict guardrails. Eastman points out that while the AI conference circuit is saturated with the buzzword "guardrails," almost everyone is exclusively referring to **LLM harness guardrails**—the model-specific safety filters and prompt boundaries popularized by vendors like NVIDIA. While essential for language models, Siemens approaches industrial autonomy through two additional, foundational layers: **coordination guardrails** and **operational guardrails**.
Coordination guardrails govern multi-actor interactions. Real-world factories are not greenfield software environments; they are populated by human operators, traditional deterministic logic, and diverse machine learning models ranging from reinforcement learning to self-driving vision algorithms. These actors must be constantly aware of each other’s concurrent actions. Operational guardrails, meanwhile, map directly to the hierarchical state machines that underpin any manufacturing facility. Whether a system is running an LLM agent, a reinforcement learning controller, or legacy deterministic rules, every model must evaluate actions against the actual physical and operational state of the plant before executing changes.
```mermaid
flowchart TD
accTitle: Siemens Three-Tier Industrial Guardrail Architecture
subgraph OperationalLayer["1. Operational Guardrails (Physics & State Machines)"]
State["Hierarchical State Machine Awareness"]
end
subgraph CoordinationLayer["2. Coordination Guardrails (Multi-Actor Sync)"]
Actors["Multi-Actor Coordination (RL, Rules, Humans, Agents)"]
end
subgraph HarnessLayer["3. LLM Harness Guardrails (Model-Level)"]
Harness["Model-Specific Math & Prompt Boundaries"]
end
Harness --> Actors
Actors --> State
State --> Control["Deterministic Control Layer (PLCs & Machines)"]
```
[05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) The Human in the Loop: Operators, Adoption Realities, and Control Leases
Albrecht frames these guardrail tiers as the core principles of Siemens' **Smart Production Orchestration (SPO)** solution. Eastman explains that the single most vital actor in any deployment is not the AI model or the engineer, but the line or unit operator. Historically, the senior operator has been the sole trusted actor authorized to modify machine setpoints because they intrinsically understand "what good looks like" and hold ultimate responsibility for hitting production targets safely. Furthermore, they are the ones who must live with the system on "day 264," long after the deployment team has packed up and left.
Eastman notes a shifting dynamic in modern manufacturing culture: management rarely forces automation down operators' throats anymore. If an operator rejects an interface within the first 30 seconds of interaction, the project is effectively dead. To bridge trust and deterministic execution, SPO relies on a strict execution hierarchy:
```mermaid
flowchart LR
accTitle: SPO Hierarchical Decision Flow to Control Layer
subgraph Governance["Governance & Validation"]
A["Model Decision"] --> B["Tailored Dynamic Envelope"]
B --> C{"Coordination Approval"}
end
subgraph Execution["Execution & Safety"]
C -->|Approved| D["Control Lease Issued (Non-Stale Time Window)"]
D --> E["PLC Setpoint Write"]
end
```
- **Dynamic Envelopes:** Tailored operating boundaries configured specifically for the active model.
- **Coordination Approval:** Verification that the model is fully aware of overall system state and concurrent actor actions.
- **Control Leases:** Time-bounded leases that define exactly how long a model's proposed decision remains fresh and valid before becoming stale and rejected.
[09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) Data Orchestration vs. State Orchestration: Laser Weeding and PLC Bottlenecks
Drawing on their shared upbringing in Kansas, Albrecht and Eastman illustrate the leap from data orchestration to **state orchestration** using an agricultural analogy. Traditional farming operated on macro-state interventions: a farmer would inspect a field and decide to treat a crop once or twice a season. Modern smart agriculture deploys autonomous vehicles nightly that identify individual weeds and zap them with targeted lasers in real time.
Industrial manufacturing is undergoing the exact same transition. A human operator typically adjusts a setpoint at a macro level—perhaps once or twice across an entire shift. An AI model, by contrast, can micro-adjust setpoints once every second to optimize efficiency in ways macro swings never could. However, this high-frequency intervention introduces entirely new failure modes:
- **Settling Time Violations:** AI models frequently attempt to write setpoint changes without waiting for physical systems to settle, firing thousands of writes within minutes and locking up the CPU of a programmable logic controller (PLC).
- **Network Bottlenecks:** High-frequency control rapidly exposes deferred infrastructure debt; legacy operational technology (OT) networks that were neglected over the prior decade immediately burn out under AI-driven write loads.
[13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) Industrial Governance at the Orchestration Layer: Physics Meets Safety
Governance in an operational technology (OT) environment differs fundamentally from corporate IT governance. While enterprise IT governance focuses heavily on cybersecurity, data privacy, and regulatory compliance, OT governance operates at the direct intersection of technology and physical reality. In plants filled with massive, high-speed rotating equipment where personnel safety is on the line, governance must be strictly deterministic.
Siemens implements this governance directly within the orchestration layer, architecturally positioned between supervisory control (SCADA/DCS) and the real-time control layer (PLCs). Eastman emphasizes that connecting AI through traditional SCADA layers introduces unacceptable latency for closed-loop setpoint adjustments. Instead, the orchestration layer sits below SCADA to enforce latency budgets, validate system state, and ensure safety interlocks before control commands reach physical hardware.
[15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) Closing the Loop: Feedback Architectures and the "AI Factory"
Albrecht and Eastman outline the transition from open-loop industrial pilots to genuine closed-loop operations. A complete industrial closed-loop architecture consists of three interconnected cycles:
1. **The Real-Time Control Loop:** The model actively writing validated setpoint adjustments directly back to the control system.
2. **The Data-to-Training Feedback Loop:** Capturing telemetry and edge decisions to continuously refine and retrain upstream models.
3. **The Drift Management Loop:** Monitoring model performance over time to detect, mitigate, and govern real-world operational drift.
Eastman compares this architecture to the autonomous vehicle industry. Companies like Tesla and Waymo pulled ahead in self-driving not merely by building better base algorithms, but by engineering closed-loop "AI factories." In these architectures, every mile driven feeds the retraining pipeline. Whoever establishes the infrastructure to log operational feedback and safely iterate models against real-world execution wins the operational race.
[17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) AI as an Organizational Capability: De-hyping LLMs on the Factory Floor
In their closing remarks, the speakers caution against treating "AI" as a standalone initiative or a universal cure-all. Many manufacturing challenges do not require machine learning at all and are far better solved with deterministic logic or traditional engineering.
When large language models are deployed, Eastman stresses that they must never be treated as reasoning engines: "They don't reason. That's not what they're doing." LLMs do not understand causality in the way humans do. Instead, their true value in industrial environments is serving as a translation layer—translating complex operational states and telemetry into natural language for operators, and parsing human operator intent back into structured system parameters. Albrecht concludes by inviting attendees to visit Siemens at Booth 831 to continue discussions on Smart Production Orchestration.
> [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) Smart Production Orchestration (SPO) — Bryan Albrecht
> [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) what good looks like — Caleb Eastman
> [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) day 264 — Caleb Eastman
> [09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) state orchestration — Caleb Eastman
> [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) AI factories — Caleb Eastman
> [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) They don't reason. That's not what they're doing. — Caleb Eastman
## Mentioned
Bryan Albrecht, Caleb Eastman, Siemens, Siemens Digital Industries, Smart Manufacturing Digital Thread & Incubation and Validation Deployment (SMDT & IVD), Smart Production Orchestration (SPO), Ai4, NVIDIA, Tesla, Waymo, Programmable Logic Controller (PLC), SCADA, DCS, MES
## Video description
Bryan Albrecht - Director, Smart Manufacturing Digital Thread & Incubation and Validation Deployment (SMDT & IVD), Siemens Digital Industries
Caleb Eastman - Field CTO - Americas, Siemens Digital Industries
What if AI in manufacturing moved beyond insights to continuously driving decisions and actions at scale?
This session explores how agentic AI and data orchestration platforms are enabling operations that sense, decide, and act in real time—directly improving throughput, asset reliability, and efficiency. Rather than replacing existing systems, this approach unlocks value from current MES, automation, and data landscapes while creating a scalable foundation for enterprise-wide outcomes.
- How do you move from siloed pilots to governed, production-scale AI operations?
- How can AI safely act—not just recommend—while keeping humans in control?
- Where should you start to drive immediate impact, and how do you scale?
西门子增设护栏层以防AI致使工厂PLC过载
工业自主化需要在SCADA之下实现确定性状态编排，因为高频模型写入会压垮传统控制器，而语言模型又无法理解工厂的物理因果关系。
Bryan Albrecht 与 Caleb Eastman 概述了西门子通过智能生产编排（SPO）实现自主工业运营的架构蓝图。他们强调，不能仅依赖基础的 LLM harness 护栏，还必须纳入多主体协同以及基于物理特性的运行状态机护栏，以防止自动化系统违反物理机器约束。高频 AI 微调存在使传统 PLC 控制器和网络不堪重负的风险，因此需要在 SCADA 之下设立确定性治理层并签发有时限的控制租约。最终，可持续的规模化应用需要借鉴自动驾驶架构的闭环反馈系统，将 LLM 视为面向操作人员的结构化翻译接口，而非因果推理引擎。
工业自主化需要三层安全护栏架构，除了模型级的 LLM 护栏外，还必须引入多角色协同护栏以及与工厂分层物理状态机紧密绑定的运行护栏。 (1:20)
一线操作员的信任是自主化落地的最大瓶颈，西门子智能生产编排方案通过动态包络与时间受限的控制租约，防止失效的模型决策被写入 PLC 执行。 (5:04)
工业控制正从宏观调节转向高频微调，但若在设置点调节中忽视机器物理稳定时间，容易造成 PLC 处理器饱和并拖垮老旧 OT 网络基础设施。 (9:30)
OT 治理直接关乎物理安全与确定性设备边界，为了避免传统 SCADA 监控层带来的延迟瓶颈，编排层必须下沉至 SCADA 系统之下以保障闭环控制。 (13:16)
Reading summary
Deploying agentic AI in industrial manufacturing requires moving beyond standard language-model safeguards to multi-tiered operational guardrails that respect physical machine dynamics, legacy networks, and human operators, according to Siemens Digital Industries' Bryan Albrecht and Caleb Eastman.
Three-Tiered Guardrails for Industrial Autonomy (1:20)
Eastman notes that industry discussions around AI guardrails focus almost exclusively on model-level LLM harness guardrails, such as prompt boundaries and safety filters popularized by vendors like NVIDIA. While necessary, Siemens argues this approach is insufficient for factories, which require coordination guardrails and operational guardrails to function safely. Coordination guardrails synchronize multi-actor environments containing human operators, deterministic logic, reinforcement learning, and vision models, ensuring each actor remains aware of concurrent activities.
Operator Trust and Time-Bounded Execution (5:04)
Within Siemens' Smart Production Orchestration (SPO) architecture, the senior line or unit operator remains the most critical actor, as they hold ultimate responsibility for safety and production targets. Modern plant culture avoids forcing tools on staff, meaning a solution will fail if an operator rejects an interface within the initial 30 seconds of interaction. To maintain safety and foster adoption, SPO uses dynamic envelopes tailored to active models, coordination approvals, and time-bounded control leases that discard model outputs if they become stale before executing a programmable logic controller (PLC) setpoint write.
Transitioning from Macro Adjustments to State Orchestration (9:30)
Industrial facilities are shifting from human-driven macro adjustments—such as manual setpoint tweaks once or twice per shift—to micro-adjustments executed every second by AI models. However, high-frequency writes introduce physical failure modes, such as violating physical settling times, which can overwhelm a PLC's central processing unit. Unregulated write loads also trigger network bottlenecks that rapidly exhaust deferred infrastructure debt across neglected operational technology (OT) networks.
Deterministic OT Governance and Latency Budgets (13:16)
Unlike enterprise IT governance centered on cybersecurity and data privacy, OT governance directly protects physical personnel and heavy rotating machinery through deterministic rules. Siemens positions its orchestration layer below supervisory control (SCADA/DCS) and directly above the real-time PLC control layer. Eastman explains that routing closed-loop setpoint adjustments through SCADA introduces unacceptable latency, whereas an intermediary orchestration layer enforces latency budgets and verifies safety interlocks before commands reach physical hardware.
Closed-Loop Feedback and Practical LLM Utility (15:32)
True autonomy mirrors the autonomous vehicle sector by building closed-loop feedback systems consisting of real-time control, data-to-training pipelines, and drift management loops that continuously log operational outcomes. Eastman cautions against treating AI as a universal fix or viewing LLMs as reasoning engines, stating plainly that LLMs lack causal understanding. On the factory floor, their practical value lies in acting as a translation layer between operational telemetry and natural language for operators.
Realizing industrial autonomy depends less on algorithm sophistication than on building deterministic, low-latency orchestration layers that enforce physical state boundaries and support operators on day-to-day operations long after initial deployment.
精炼阅读摘要
Siemens Digital Industries 的 Bryan Albrecht 与 Caleb Eastman 指出，工业自主化不能仅依赖通用大语言模型（LLM）的安全护栏，必须在可编程逻辑控制器（PLC）控制层之上构建结合物理状态机、多行动主体协同与时间租约的确定性编排层。
超越模型护栏的三层工业防护体系 (1:20)
Eastman 提出，AI 行业通常讨论的“护栏”多局限于以 NVIDIA 为代表的模型线束护栏（LLM Harness Guardrails），即模型数学边界与提示词过滤。但在实际工厂环境中，绿地软件假设并不成立，生产现场同时存在一线操作员、传统确定性逻辑、强化学习控制器以及计算机视觉算法等多类行动主体。因此，工业自主化必须建立三层防护体系，在线束护栏之上增加协同护栏（Coordination Guardrails）与运行护栏（Operational Guardrails）。
一线操作员接纳度与动态控制租约机制 (5:04)
在 Siemens 的智能生产编排（SPO）方案中，资深生产线或单元操作员是决定项目成败的核心角色。Eastman 强调，现代制造企业管理层已极少强行推行自动化系统；若操作员在接触界面的前 30 秒内产生抵触，项目通常难以推进。系统上线后长期平稳运行的关键在于让操作员保留最终掌控感，因为操作员了解正常生产的真实标准并承担最终安全责任。
为将模型决策安全地转化为底层执行，SPO 构建了分层决策流。系统针对具体模型配置动态包络（Dynamic Envelopes），在验证全局系统状态与多主体动作无冲突后签发协调许可。随后，系统发放具有明确有效时间窗口的控制租约（Control Lease）；一旦模型提议的设定值在时限内未被执行或发生过时，指令将直接失效并被丢弃，从而防止过时决策写入可编程逻辑控制器（PLC）。
高频状态编排暴露的底层控制与网络瓶颈 (9:30)
Albrecht 与 Eastman 指出，工业场景正从低频的数据编排转向高频的状态编排。传统模式下，操作员一个班次仅对设备设定值进行一至两次宏观调整；而 AI 模型能够以每秒一次的微调频率优化运行效率。然而，这种微观高频干预带来了全新的物理与基础设施冲突。
高频写入极易引发物理沉降时间违规：若 AI 模型在机械系统尚未完成物理稳定时持续下发指令，短时间内涌入的数千次写请求会导致 PLC 的 CPU 发生死锁。此外，这种高频交互暴露了工厂长期拖欠的基础设施技术债务，导致过去十年缺乏维护升级的陈旧操作技术（OT）网络因过载而瘫痪。
编排层治理架构与类自动驾驶闭环演进 (13:16)
Eastman 阐明，OT 环境的治理直接作用于高速重型旋转设备与人身安全，具有与侧重数据隐私、网络安全的企业 IT 治理完全不同的确定性要求。为规避传统 SCADA/DCS 系统引入的高延迟，Siemens 将编排治理层置于 SCADA 之下、PLC 实时控制层之上，直接强制执行延迟预算、校验系统状态并确保物理安全联锁生效。
实现真正工业闭环需要三套相互关联的循环架构：向控制系统写入设定值的实时控制回路、采集遥测与边缘决策以迭代模型的训练反馈回路，以及监测并缓解模型漂移的治理回路。Eastman 援引特斯拉与 Waymo 在自动驾驶领域的路径说明，决定工业竞争优势的不仅是基础算法本身，而是能够记录真实运行反馈并安全迭代模型的“AI 工厂”闭环工程能力。
Eastman 与 Albrecht 强调，制造领域不应将 AI 视作独立目标或通用灵药，许多工厂问题通过确定性工程逻辑解决更为有效；LLM 缺乏因果推理能力，其在车间现场的核心定位应是双向翻译工具，负责将复杂的系统遥测与运行状态解析为自然语言，并将操作员意图转化为受控的结构化系统参数。
Detailed notes — zh
## Background
该视频由行业会议频道 Ai4 于 2026 年 8 月 23 日发布在 YouTube 上。演讲由西门子数字工业软件（Siemens Digital Industries）的部署总监 Bryan Albrecht 与美洲区现场 CTO Caleb Eastman 共同主讲，主题聚焦于如何利用 Agentic AI 与数据编排平台推动工业制造走向全流程自主闭环，在不推翻现有 MES 与自动化架构的前提下安全、可控地实现生产规模化提效。
## TL;DR
演讲深入剖析了工业制造从单点 AI 试点迈向自主闭环控制的工程实践与落地挑战。西门子团队提出了涵盖大模型束缚、多智能体协同与底层状态机运行的三层安全护栏体系，并通过控制租约与模型包络确保一线操作员始终掌握核心控制权。同时，演讲者厘清了数据编排与状态编排的本质差异，指出了 AI 高频微调容易引发的 PLC 算力锁死及老旧 OT 网络拥堵等物理瓶颈，强调必须将治理层下沉至 SCADA 与 PLC 之间以严守物理安全底线，最终通过构建类似自动驾驶的“三大闭环”实现工业 AI 工厂的高频迭代，并将大模型严格定位为交互翻译层而非核心推理引擎。
## Atmosphere & style
整场演讲充满务实干练的工业一线工程气息，两位主讲人配合默契且风趣幽默，常用“堪萨斯激光除草”、“车间前30秒生死线”等生动比喻打破枯燥的架构讲解。现场交流直面物理碰撞、PLC 崩溃与网络瘫痪等真实痛点，毫不留情地刺破纯算法层面的浮夸炒作，给观众带来一种来自车间现场、兼具技术前瞻性与深厚工业敬畏感的清醒沉浸感。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=4s) 工业自主控制超越了传统数据编排，核心在于建立紧密结合物理上下文的三层安全护栏体系：LLM 束缚护栏、多智能体协同护栏与复杂状态机运行护栏。
- [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) 车间新系统的落地成败取决于一线操作员的接受度，高级操作员是掌握最佳工艺设定点并负责长期运维的核心主体。
- [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) AI 决策下发至底层控制层必须经过严格的分层机制，包括模型运行包络验证、多主体协同审批以及防止过期写入的控制租约授权。
- [09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) 工业智能化正从人类操作员的低频宏观调优演进为 AI 智能体的高频微调，但高频控制必须严格感知物理系统的响应与稳定时间，防止写入风暴瘫痪系统。
- [12:25](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=745s) AI 高频指令下发直接暴露了底层基础设施的瓶颈，容易引发 PLC 算力锁死，并对长期缺乏升级改造的 OT 网络带宽构成严峻考验。
- [13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) 工业编排层必须打破传统层级下沉至 SCADA 与底层 PLC 之间，以消除通信延迟，在最靠近物理设备的位置执行严格的准入判定与防损规则。
- [13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) OT 领域的 AI 治理直接面对运转的重型机械与物理规律，必须在常规企业 IT 数据合规与网络安全之上，强制叠加确定性的物理安全防护。
- [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) 工业制造的终极目标是构建包含底层指令写回、数据自动再训练与模型漂移治理三大相互嵌套的闭环系统。
- [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) 工业 AI 竞争借鉴了自动驾驶 AI 工厂模式，其核心在于打通真实工况数据闭环，依靠累积运行里程实现算法的高频自我迭代。
- [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) AI 智能体应被视为企业的组织与运营能力而非孤立技术项目，其应用必须紧密围绕具体业务与工程痛点展开。
- [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) 大语言模型（LLM）不具备因果推理能力，在严谨的工业闭环控制中应严格定位为连接人类操作员与复杂系统的双向翻译和接口层。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=4s) 工业制造中的自主系统愿景与三大安全护栏体系
西门子数字工业软件（Siemens Digital Industries）的部署团队总监 Bryan Albrecht 与美洲区现场首席技术官（Field CTO - Americas）Caleb Eastman 共同开场。Bryan 指出，在过去六个月中，制造业迈向自主化成为了工业界最炙手可热的话题之一。Caleb 则幽默地界定自己的职责范围是“大西洋这一侧”：只要客户希望在工厂车间实现某种形式的自主化（Autonomy），他和 Bryan 的团队就会第一时间介入现场。演讲伊始，Bryan 现场调查了有多少与会者正在从事面向制造业全自主系统的项目，台下举手者众，印证了工业界对这一方向的迫切探索。
Bryan 指出，行业目前已经超越了单纯的“数据编排”（Data Orchestration）需求，普遍意识到必须赋予数据工业“上下文”（Context），但更关键的核心在于建立与上下文紧密相连的安全护栏（Guardrails）。Caleb 强调，虽然在当前的各类 AI 大会上，英伟达等厂商提及的护栏几乎全是指“大语言模型束缚护栏”（LLM Harness Guardrails，即输入输出过滤与提示词安全），但在工业自主控制场景下，西门子提出了独有的三层护栏体系：
```mermaid
flowchart TD
accTitle: 工业自主控制的三层安全护栏体系
subgraph Guardrails["工业自主控制安全护栏 (Industrial Autonomy Guardrails)"]
G1["1. LLM 束缚护栏 (LLM Harness Guardrails)<br/>特定针对语言模型的输入/输出过滤与防幻觉"]
G2["2. 协同护栏 (Coordination Guardrails)<br/>协调人类操作员、RL强化学习、自动驾驶算法等多智能体"]
G3["3. 运行护栏 (Operational Guardrails)<br/>工业控制复杂状态机（State Machine）层级映射与统一通信"]
end
G3 --> G2 --> G1
```
- **LLM 束缚护栏（LLM Harness Guardrails）**：行业中最常见、专门针对 LLM 类模型设计的传统安全网，主要防范生成内容偏离或幻觉。
- **协同护栏（Coordination Guardrails）**：在通往完全自主的过渡期，现场必然同时存在人类操作员、强化学习（RL）模型、类自动驾驶算法等多种执行主体（Actors）。协同护栏用于确保各模型输出在多智能体共存的环境中互相感知、协调动作。
- **运行护栏（Operational Guardrails）**：任何制造工厂的控制系统本质上都是极其复杂的“分层状态机”（Hierarchical State Machines）。无论是大语言模型、规则系统还是经典数学控制模型，都必须在统一的工业状态感知下运行，防止破坏物理底层逻辑。
---
### [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) 核心参与者：一线操作员的决定性角色与控制决策层级
这套护栏原则依托于西门子的“智能生产编排”（Smart Production Orchestration，简称 SPO）解决方案。Bryan 与 Caleb 深入拆解了系统中的参与者角色。尽管系统上层存在各种 AI 智能体和机器，但车间里最核心的执行者始终是产线操作员（Line Operator / Unit Operator）。在传统工厂中，高级操作员（Senior Operator）是唯一被信任去调整设备设定点（Set Point）的人，他们清楚“优秀的生产状态到底是什么样”，更重要的是——当部署团队在“第 264 天”撤场交付后，真正要日复一日面对并维护这套系统的人正是这些一线员工。
Caleb 尖锐地指出，车间新系统的生死往往决定于一线工人接触系统的“最初 30 秒”。如果操作员第一印象排斥该系统，项目基本上就宣告死亡；如今的管理层也越来越倾向于将决策权下放：“操作员喜欢它，我们就喜欢；操作员讨厌它，我们就不干。”为了确保 AI 能安全、受控地向控制层下发决策，SPO 构建了一套严格的分层下发机制：
```mermaid
flowchart LR
accTitle: AI 决策下发至控制层的分层验证流程
A["AI / 模型决策输出"] --> B["模型包络验证 (Model Envelope)"]
B --> C["协同审批 (Coordination Approval)"]
C --> D["控制租约授权 (Control Lease)"]
D --> E["写入 PLC / 底层控制层"]
```
- **模型包络（Model-tailored Envelopes）**：针对特定模型类型量身定制的物理运行安全边界，限制其参数调整范围。
- **协同审批（Coordination Approval）**：确认模型在做出调整指令时，是否已知晓系统中其他正在发生的动作与状态。
- **控制租约（Control Lease）**：赋予模型一个极具时效性的写入窗口期，防止因通信延迟或系统变动导致“陈旧过期的决策”（Stale Decisions）被错误写入底层控制硬件。
---
### [09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) 数据编排 vs. 状态编排：从“农业激光除草”看微观实时调优
Bryan 和 Caleb 转向了目前工业智能化讨论的核心议题：**数据编排（Data Orchestration）与状态编排（State Orchestration）的本质区别**。由于两人都来自堪萨斯州（Kansas），他们借用美国农业带的演进生动打比方：传统农业中，农户通常根据宏观季节状态在一季中只对农田施肥打药一到两次；而现代智慧农业中，自主农机每晚穿行于田间，利用视觉算法动态识别杂草并实时用激光精准将其消灭。
这种“宏观定点”到“高频微调”的转变正在制造业真实上演：
- **人类操作员的宏观控制**：工人在一个班次（Shift）运行期间，通常只针对设定点（Set Point）进行一到两次宏观层面的经验调整。
- **AI 智能体的微观控制**：AI 可以在运行中每秒对设定点进行一次细微微调（Micro-adjustments），实现人类粗放调整根本无法达到的工艺极值与能效优化。
然而，微观调控也带来了巨大的工程挑战。Caleb 警告，传统的工业控制流程最终会汇聚到仪表盘供人查看并反应，而自动化写入则完全不同。过去曾出现过严重的现场故障：由于 AI 模型缺乏对物理系统“稳定时间”（Settle Out Time）的感知，在物理对象还未完成响应前就不断重试，短短几分钟内向底层控制系统发起了数千次写入请求（Write Storm），直接导致控制系统剧烈震荡甚至瘫痪。因此，精准把握状态编排、控制动作节奏与物理反馈，是自主 AI 能够真正落地的关键前提。
### [12:25](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=745s) 控制系统的瓶颈：PLC 算力过载与 OT 网络的严峻考验
将 AI 直接部署到控制系统中时，系统暴露出的第一批故障模式往往是传统自动化时期未曾见过的。在以往的工业架构中，没有任何上层应用能够以 AI 的高频速度向底层下发指令，而一旦 AI 以极高的吞吐和速率向可编程逻辑控制器（PLC）发出控制请求，便可能直接导致“PLC 的 CPU 彻底锁死”。这并不是软件算法本身崩溃，而是底层控制硬件的物理算力被瞬间耗尽。
现场实施中最残酷的现实在于：一旦将 AI 推入实时控制链路，整个系统的瓶颈会立刻无所遁形。Caleb Eastman 带着调侃指出，最先亮起红灯的往往是客户“过去十年里觉得不花钱也无所谓的领域”。企业总以为现有的网络带宽或老旧通信架构足够支撑运转，但实践证明，运营技术（OT）网络是首先被高负荷数据流击垮的两大核心瓶颈之一。许多制造企业不得不为了承载 AI 控制流而彻底改造升级其底层网络。
- **新型失效模式**：高频下发控制指令导致 PLC CPU 算力超载锁死。
- **首要基础设施瓶颈**：长期缺乏升级改造的 OT 现场网络体系。
---
### [13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) 工业治理模型：编排层下沉与“技术碰撞物理”的安全底线
高频交互带来的系统脆弱性自然引出了工业现场最关切的核心疑问：“我们如何信任这一套系统？”针对制药（Pharma）以及食品与饮料（Food & Beverage）等强监管行业的差异化合规诉求，西门子将治理模型（Governance）嵌入到了系统的“编排层”（Orchestration Layer）。Caleb Eastman 强调，这套编排机制绝不能悬浮在传统的 SCADA 监控层之上，而是必须下沉到 SCADA 与底层 PLC 之间。传统 SCADA 或 DCS（分布式控制系统）网络连接会引入过高的通信延迟，AI 若要进行毫秒级的安全判定与干预，就必须在更靠近物理设备的最下层直接执行规则准入与指令过滤。
工业场景中的治理体系与纯企业级 IT 治理有着本质的区别。IT 治理的核心主要聚焦于网络安全（Cyber Security）与法规合规（Regulatory Concerns）；但在 OT 领域，治理必须直面物理世界的法则。Caleb 形容这里是“技术与物理规律直接碰撞的地方”，现场不仅运转着巨大的旋转重型机械，任何一个误动作都极易导致人员伤亡。因此，OT 环境下的 AI 治理是在 IT 治理规则之上，额外叠加了一整套硬性的物理安全防护栏与确定性规则。
```mermaid
flowchart TD
accTitle: Industrial AI Governance and Orchestration Architecture
subgraph Enterprise["IT 与企业管理层"]
A["IT 治理与合规模型<br/>（网络安全 / 行业法规）"]
end
subgraph Orchestration["西门子 SPO 编排与治理层"]
B["编排层 Orchestration Layer<br/>（下沉至 SCADA 与 PLC 之间）"]
C["物理安全规则与准入护栏<br/>（防碰撞 / 防过载 / 延迟控制）"]
end
subgraph Physical["OT 与物理执行层"]
D["SCADA / DCS 监控系统"]
E["PLC 控制器与现场设备"]
F["重型旋转机械与产线工人"]
end
A --> B
D -. 旁路监控.-> B
B --> C
C ==> E
E --> F
```
- **编排层架构位置**：打破传统层级，下沉至 SCADA 与 PLC 控制器之间以规避网络延迟。
- **OT 治理双重属性**：在常规企业数据合规与网络安全之上，强制绑定物理安全（Safety）与机械运动防损规则。
---
### [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) 从开环到闭环：制造系统的三大闭环与“AI 工厂”迭代逻辑
从设备上下文、安全护栏、智能体角色划分到状态编排与治理，所有的架构设计最终都指向制造企业的终极目标——实现生产全流程的“闭环系统”（Closed-Loop System）。工业场景中的闭环并非单一的数据往返，而是由三个相互嵌套的独立回路构成的有机整体：
1. **控制执行回路**：AI 模型直接向底层控制系统执行写回操作（Write-back）。
2. **数据反馈与训练回路**：产线运行数据实时回流，驱动新模型的自动化训练与迭代。
3. **漂移监控与治理回路**：长期追踪并动态校正模型在物理工况变化下的模型漂移（Model Drift）。
Caleb 将工业闭环系统的演进路径与自动驾驶汽车领域的突破进行了类比。自动驾驶技术之所以能从概念走向成熟，核心就在于特斯拉（Tesla）和 Waymo 等领头羊率先打通了模型研发与现场反馈的闭环，构建出了所谓的“AI 工厂”（AI Factories）。在完善的基础设施建立之后，行业竞争便转变成了“数据里程累积”的较量——谁在真实系统中为模型跑出的里程数最多，谁就能实现最高频的自我迭代并保持领先。工业制造系统正在经历完全相同的范式转移。
- **闭环系统的三大支柱**：控制系统指令写回、数据反哺模型再训练、长效模型漂移治理。
- **自动驾驶启示**：工业 AI 竞争的本质是跑通“数据闭环与自我迭代”，通过真实工况运行里程持续提升系统智能化水平。
---
### [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) AI 智能体作为组织能力：大模型的翻译本质与理性应用
随着演讲进入尾声，Bryan Albrecht 抛出了核心设计原则：必须将 AI 智能体视为一种“组织或运营能力”（Organizational Capability），而非仅仅是一个孤立的技术工具。企业如果仅仅将其立项为某个单一的“AI 项目”，往往会在落地时陷入误区。真正的起点必须回归到业务本质：究竟要解决什么具体的工程与运营痛点。
Caleb 对当前大语言模型（LLM）的热潮给出了务实的冷思考。他指出，制造现场的大多数核心问题根本不需要 LLM 来解决；即便在部分场景中引入了 LLM，也绝不能将其用作系统的“核心推理引擎”。因为大模型本质上“根本不具备推理能力”（"They don't reason. That's not what they're doing."），它们既无法真正理解因果关系，更不能替代物理控制逻辑。在工业闭环架构中，LLM 最恰当的定位仅仅是充当复杂工业系统与人类操作员之间的“双向翻译层”。Bryan 在演讲结束时向现场观众致谢，并邀请大家前往 831 号西门子展台，共同深入探讨如何借助西门子 Smart Plant Operations（SPO）平台落地可信的智能体工业运维。
- **技术定位**：切勿为了 AI 而搞“AI 项目”，优先评估业务痛点是否真正依赖模型技术。
- **LLM 的工业角色**：大模型不具备因果推理能力，在严谨的工业控制中应严格定位为人机交互的“翻译与接口层”。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=4s) 只要客户希望在工厂车间实现某种形式的自主化，我和 Bryan 的团队就会第一时间介入现场。 — Caleb Eastman
> [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) 车间新系统的生死往往决定于一线工人接触系统的最初 30 秒。 — Caleb Eastman
> [05:04](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=304s) 操作员喜欢它，我们就喜欢；操作员讨厌它，我们就不干。 — Caleb Eastman
> [09:30](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=570s) AI 可以在运行中每秒对设定点进行一次细微微调，实现人类粗放调整根本无法达到的工艺极值与能效优化。 — Caleb Eastman
> [12:25](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=745s) 一旦将 AI 推入实时控制链路，最先亮起红灯的往往是客户过去十年里觉得不花钱也无所谓的领域。 — Caleb Eastman
> [13:16](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=796s) 这里是技术与物理规律直接碰撞的地方，现场运转着巨大的旋转重型机械，任何误动作都极易导致人员伤亡。 — Caleb Eastman
> [15:32](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=932s) 谁在真实系统中为模型跑出的里程数最多，谁就能实现最高频的自我迭代并保持领先。 — Caleb Eastman
> [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) 必须将 AI 智能体视为一种组织或运营能力，而非仅仅是一个孤立的技术工具。 — Bryan Albrecht
> [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) They don't reason. That's not what they're doing. — Caleb Eastman
> [17:47](https://www.youtube.com/watch?v=9C2iDUh8PWk&t=1067s) 在工业闭环架构中，LLM 最恰当的定位仅仅是充当复杂工业系统与人类操作员之间的双向翻译层。 — Caleb Eastman
## Mentioned
Bryan Albrecht, Caleb Eastman, 西门子数字工业软件 (Siemens Digital Industries), 智能生产编排 (Smart Production Orchestration / SPO), 英伟达 (NVIDIA), 特斯拉 (Tesla), Waymo, 可编程逻辑控制器 (PLC), SCADA 监控系统, 分布式控制系统 (DCS), 大语言模型 (LLM)
## Video description
Bryan Albrecht - Director, Smart Manufacturing Digital Thread & Incubation and Validation Deployment (SMDT & IVD), Siemens Digital Industries
Caleb Eastman - Field CTO - Americas, Siemens Digital Industries
What if AI in manufacturing moved beyond insights to continuously driving decisions and actions at scale?
This session explores how agentic AI and data orchestration platforms are enabling operations that sense, decide, and act in real time—directly improving throughput, asset reliability, and efficiency. Rather than replacing existing systems, this approach unlocks value from current MES, automation, and data landscapes while creating a scalable foundation for enterprise-wide outcomes.
- How do you move from siloed pilots to governed, production-scale AI operations?
- How can AI safely act—not just recommend—while keeping humans in control?
- Where should you start to drive immediate impact, and how do you scale?
[Source](https://www.youtube.com/watch?v=9C2iDUh8PWk)
Source quotes
[0:04] Caleb Eastman: 只要客户希望在工厂车间实现某种形式的自主化，我和 Bryan 的团队就会第一时间介入现场。
[5:04] Caleb Eastman: 车间新系统的生死往往决定于一线工人接触系统的最初 30 秒。
[5:04] Caleb Eastman: 操作员喜欢它，我们就喜欢；操作员讨厌它，我们就不干。
[9:30] Caleb Eastman: AI 可以在运行中每秒对设定点进行一次细微微调，实现人类粗放调整根本无法达到的工艺极值与能效优化。
[12:25] Caleb Eastman: 一旦将 AI 推入实时控制链路，最先亮起红灯的往往是客户过去十年里觉得不花钱也无所谓的领域。
[13:16] Caleb Eastman: 这里是技术与物理规律直接碰撞的地方，现场运转着巨大的旋转重型机械，任何误动作都极易导致人员伤亡。
[15:32] Caleb Eastman: 谁在真实系统中为模型跑出的里程数最多，谁就能实现最高频的自我迭代并保持领先。
[17:47] Bryan Albrecht: 必须将 AI 智能体视为一种组织或运营能力，而非仅仅是一个孤立的技术工具。
[17:47] Caleb Eastman: They don't reason. That's not what they're doing.
[17:47] Caleb Eastman: 在工业闭环架构中，LLM 最恰当的定位仅仅是充当复杂工业系统与人类操作员之间的双向翻译层。
Source quotes
[1:20] Caleb Eastman: LLM harness guardrails
[1:20] Caleb Eastman: coordination guardrails
[1:20] Caleb Eastman: operational guardrails
[5:04] Bryan Albrecht: Smart Production Orchestration (SPO)
[5:04] Caleb Eastman: what good looks like
[5:04] Caleb Eastman: day 264
[9:30] Caleb Eastman: state orchestration
[15:32] Caleb Eastman: AI factories
[17:47] Caleb Eastman: They don't reason. That's not what they're doing.
Original transcript
[0:04] Hello. Thank you all for being here. Um, as uh as
we were introduced, my name is Brian Alrech. I'm a
[0:12] director for uh deployment teams here at Seammens.
And uh today we're going to be talking about one
[0:20] of the subjects that has been particularly
popular in the last six months for us. Uh,
[0:26] but before we get started, I'm going to let
Caleb talk a little bit more about himself
[0:30] and what he does for SEM as well. Oh, hi. So, I'm
I'm Caleb. I work quite closely with Brian. Um,
[0:37] I'm the field CTO Americas, uh, which is to say
this side of the Atlantic. Um, and uh, and what
[0:45] I really do at Seammens is generally speaking,
if someone is wanting to do something involving
[0:53] autonomy of some sort um, on their manufacturing
floor, then that's generally when I get involved
[0:58] and when Brian Steam gets involved. All right. So,
before we get started within our our presentation,
[1:05] we just wanted to get a show of hands. How many
of you are involved uh with something that's going
[1:12] towards fully fully autonomous systems within
manufacturing? Okay, good. Good amount of you.
[1:20] Okay. So, the the first place that we like to
start is, you know, we've all been in a place
[1:28] where the the customer or whoever is asking
for data orchestration, but now we're starting
[1:36] to understand and know that we need to have
context behind that. But what we really want
[1:42] to kind of show behind that is the guardrails and
the relationship with the context. So, Caleb, can
[1:48] you talk to us a little bit about your guardrails
that you have up here? Yeah, so you're going to
[1:52] probably at this conference, you're probably going
to have the word guard rails thrown at you a lot.
[1:57] Um, and when we at Seammens talk about guardrails,
we're uh usually you're guardrails are oftentimes
[2:04] used in the context of LLMs in particular. Um,
and that's a type of guardrails specific to that
[2:11] type of model that pretty much everybody who
you talk who you hear the word guardrails from,
[2:17] they'll be talking about this type which is in
column one here. Uh, which is the tradition.
[2:23] Those are LLM harness guardrails. So you might if
you hear Nvidia talk about guardrails or whatever,
[2:28] this is what they're referring to and it's
specific to LLMs. So um and because LLMs are are
[2:35] uh are now a fairly big player in AI uh in fact is
probably the number one thing being talked about
[2:41] this conference today then um it's important to
understand them in the context of everything else
[2:47] but when we talk about autonomy or or control
back to us writing control decisions back from
[2:54] a model regardless of the type of math used inside
it. Um there are two other guardrails that we at
[3:01] Seammens think about and as far as we know we're
the only people talking about it. That's fine.
[3:07] Um is what we call coordination and operational
guardrails. And coordination guardrails come into
[3:14] play whenever you're dealing with our timer is not
going. Oh well. Um we still have 20 minutes. Uh
[3:22] uh coordination guardrails you know if you're on
a path towards autonomy you're not there right
[3:27] now. you have operators on the floor, you have
other other types of of actors on the system,
[3:33] not just an LLM agent. And whenever you have that
those types of actors on the system, they have to
[3:39] know about each other. So coordination guard rails
are how does the outputs of one model regardless
[3:46] of the type. So this is extending beyond just
LLMs. We're talking about other types of math
[3:52] you whether it's you know reinforcement learning
or any of the other types that are relevant in
[3:57] the self-driving uh uh world. Uh those have to
be coordinated with other actors in the system.
[4:03] So there are elements to that and the way that
you go about deploying it that matter. And then
[4:09] the last is operational guardrails which is how
you how you know about the hierarchical state.
[4:16] So essentially any any any control system in any
manufacturing plant is ultimately a set of complex
[4:24] state machines. Those those state machines and the
way they interact with each other has to be has to
[4:33] be communicated universally across all models. It
doesn't matter if it's LLM or if it's another type
[4:39] of model or if it's or if it's a uh you know just
a rulesbased system. They all have to be aware of
[4:45] each other. So you have to know about the system,
what's going on right now. You have to know what
[4:51] each other are doing and then you can deal with
model specific type math uh type type guardrails.
[4:57] And so that's that's the first principle that we
tend to follow here. Yeah. So the the principles
[5:04] that are we're talking about are revolving
around a solution that we call smart production
[5:09] orchestration. Uh we at Seammens we call that SPO
in short. So this is the first principle that we
[5:15] abide by. The second one is understanding who our
actors are. And in this day and age, I think uh we
[5:24] have a lot of actors. And Caleb's going to talk to
us a little bit about who the actors are as within
[5:31] our deployment teams, who we're seeing as actors,
whether that's people, machines, AI agents sitting
[5:38] on top, uh, and what their relationship is within
SPO. [snorts] So, I'm going to change the slide
[5:46] here. That's fine. We'll probably stay on this
one for the rest of time. Actually, the next one.
[5:53] Okay. So um there are uh you know I would say the
primary actor or the primary person that has to
[6:04] in some way know what the what the intent of the
model is know how to turn it off know what's going
[6:11] on with it what's attempting to do is generally
speaking the operator of the machine. Um, and
[6:18] and here we're talking about the line operator or
the unit operator. Depends on what type of process
[6:24] control or factory control you do, what this
person's name is or what this person's title is,
[6:29] but generally speaking, it's going to be
the person who's responsible ultimately
[6:32] or historically been responsible for the the the
output of that day being met safely. Basically,
[6:40] that operator is by far the most important
person in in in when you're dealing with
[6:46] um manufacturing systems because they are the one
who knows what good looks like generally speaking,
[6:53] especially a senior operator. And they're also the
person who has to live [snorts] with this after
[6:59] all your after the deployment team leaves on day
264. This is the person that has to live with it.
[7:06] And so we and and historically this has
been the only actor on the system. There
[7:11] hasn't been that you generally speaking if a
senior operator wants to adjust a set point on
[7:15] a machine that's the person who you would trust
to do that. Um so so when that person now has to
[7:24] um interact with other types of with with new
new types of systems um their adoption I mean
[7:32] generally speaking you might have and depending
upon the manufacturing facility 30 seconds on
[7:37] that first day of them interacting with it where
they will either want to do it or not do it and if
[7:42] they don't want to do it the project is probably
dead un you know unless management gets involved
[7:48] And you know uh it's it's interestingly enough you
see management making less and less decisions. If
[7:55] the operator likes it, we like it. If the operator
doesn't like it, we don't like it. It's generally
[7:59] how it works nowadays. So it it the decision
is being made in operations um about whether
[8:06] or not you're going to do this. So, [snorts] and
um we have here kind of a what we like to call a
[8:13] hierarch hierarchical what is how does a decision
get driven down to the control layer basically and
[8:24] these are the different these are the different um
guard rail uh guardrails if you will that that we
[8:32] use ourselves. We use envelopes that are driven
that are tailored to the model itself. There are
[8:40] um whether there's coordination approval which
is to say is that particular model um aware what
[8:48] else is going on the system when it makes that
decision. Then we do what's called a control
[8:52] lease. Now this is a period of time that [clears
throat] we consider the decision of the model to
[8:59] be not stale and it matters um and it matters
where in the system didn't really matter to me
[9:05] if people heard me anyway. Um so then then and so
I could go into this but this is longer than a you
[9:13] know 10 minutes 20 minute talk. Yeah. So that's
this is one of those things where in order to
[9:18] really go into depth of the how and everything and
the why and you know what the lessons we learned
[9:24] that caused us to to to do it this way um you'll
have to do that after the speech probably. Yeah.
[9:30] Yeah. So Caleb you were you were talking about so
now we've talked about context and the guardrails
[9:36] and we know as we learn more of the environment
that the customer has starting to understand the
[9:42] actors within it. One thing that we've talked
about in the a lot recently is the difference
[9:49] between data orchestration and state orchestration
um within our solution. This is something that's
[9:57] probably one of the most important things to our
team is making sure we understand the state of all
[10:03] those actors to make sure we are making the right
decisions. Can you can you talk to us a little bit
[10:10] about the state orchestration because uh there's
there's one thing that uh me and Caleb always talk
[10:16] about because we're both from Kansas and so we
have a little bit of the heartland ingrained into
[10:22] us and so uh we always relate everything back to
uh farming it seems and so uh one thing that was
[10:32] interesting for for me was we were talking about
where farming came from where they still may be is
[10:41] uh you would have farmers where they would plant
a crop. They would make a decision to treat that
[10:46] crop maybe once a season, twice a season. So they
were they were treating the state of that crop.
[10:51] But this day and age, they have autonomous
vehicles that are going through nightly that
[10:58] are looking for weeds, uh taking care of those,
zapping them out with lasers, and that's acting
[11:04] every single night. And so Caleb, talk to us
a little bit about how we're seeing that in
[11:10] the manufacturing world. So yeah, so that so the
interesting thing, you know, generally speaking,
[11:16] the the places this is most effective is is if
you have a human operator who is adjusting a
[11:23] set point, they're doing it in a macro level.
They might adjust the set point once or twice
[11:28] um you know, while the product is running through
it during a shift. Um but an AI model might tweak
[11:34] the set point once a second for example and
the difference the difference in how much time
[11:40] you have to react to something being going going
wrong is is much much smaller. So the opportunity
[11:46] obviously is is that the small tweaks can can
achieve things that large macro swings cannot.
[11:52] But the problem now is that uh the you care very
much more how things um since it's not ending up
[12:01] in a dashboard for someone to look at and then
react to. Now you have AI models who actually can
[12:07] uh who you know in the past I've seen AI models
that will uh make an adjustment adjustment not
[12:14] wait long enough for the system to settle out and
try to make another adjustment. So it'll do it'll
[12:18] attempt to do like a few thousand writes to the
control system within a few couple minutes and
[12:25] lock up the CPU of the PLC. That's a thing that
can happen. Um that is a failure mode that wasn't
[12:31] understood very well prior to AI because you were
never going to have something act on it that fast.
[12:37] But the truth is is you will find your bottlenecks
very very quickly um if you're deploying AI in
[12:42] a control system. Um, and often times it's the
network or it's whatever you didn't whatever you
[12:49] thought was okay to not invest in over the last
10 years probably generally speaking. Um, whatever
[12:53] whatever you thought was fine uh is going to be
the thing that will be the first uh will be the
[12:59] thing that's a trouble first. I've [snorts] seen
a lot of people have to upgrade their network um
[13:04] considerably because they wanted to do something
like this and that's one of the two bottlenecks.
[13:09] OT networks tend to be one of the two bottlenecks
we see um get burned fast. So, so this leads us
[13:16] to probably one of the questions we get asked the
most. Uh I'm sure you all have seen this uh with
[13:22] any projects that you might be in is how do we
trust this? And so this is where our governance
[13:29] model comes into play quite often. And so how does
governance run within SPO and how do we use this
[13:40] uh if we're if we're looking at a pharma industry,
how does that relate towards something within food
[13:47] and beds? Oh, that's a good question. This is this
is governance right here. So if you're wondering
[13:53] you know what governance looks like having a
having an ordered way in which something can
[14:00] happen and making and enforcing that and generally
speaking at the what we call the orchestration
[14:07] layer which I would I will argue happens between
um SCADA and the PLCs. uh that that's one thing
[14:16] that will be a new I would say a new uh thing
is that generally speaking the connection that
[14:22] you have between skate and the PLCDs is or if
you have a DCS system is not it it it absorbs
[14:30] too much latency basically that you need you
need that time uh to do c to to to so you need
[14:37] to actually be below the ska layer if you can
be um that's governance having rules like that
[14:43] is governance Having a way that you allow this
to pass through is governance. And it looks a
[14:48] little bit differently in operational technology
than it does in enterprises because governance in
[14:52] enterprises is almost entirely around governance
and cyber security and regulatory concerns which
[15:00] um which in in operational technology you are
also adding safety to it and you're also adding
[15:07] that you know you are that's the point where
technology meets physics and that's where all your
[15:11] big rotating machines are. It's where it's where
people can get hurt pretty easily. And so we have,
[15:17] you know, so that governance in OT looks like
governance in IT and some additional rules
[15:25] that have to do with the fact that you're dealing
directly with physics. So that's the so that so we
[15:32] talked about context and guardrails relationships
there. We also talked about the actors then state
[15:39] orchestration and now governance. What is
this all leading to? where where we see
[15:44] this leading is closed loop systems uh within
manufacturing. That's often what our customers
[15:51] are looking at. And so uh within this solution,
that's that's where this is headed. So, Caleb,
[16:00] talk to us a little bit about why the closed loop
system with SPO uh how we're seeing a path forward
[16:09] with this with compared to the openloop systems
that we're currently working on. Uh so, you know,
[16:17] when we talk about closed loop, there's actually
there's there's essentially three separate loops
[16:23] within a closed loop system. One is the model
writing back to the control system. One is the
[16:31] way that you take data and then train new models
using that. That's a loop. And also there's a loop
[16:38] around how you around how you deal with things
like model drift over time and things like that.
[16:47] Um how you manage that. So when when we talk about
openloop systems the the first thing that you in
[16:55] any self-driving system is you close the model
development loop and the feedback loop. This is
[17:02] actually what um the uh self-driving cars did so
good at is they basically made a system to where
[17:09] uh the the way that they got feedback they were
able to iterate towards improvement over time and
[17:16] uh and by creating what they actually called
uh AI factories, what they called and and uh so
[17:22] what ended up happening was whoever put the most
miles on the car once the system was set up was uh
[17:29] or the the most miles on the on the model, if you
will, once the system was set up were the ones who
[17:34] won, which is why you'll notice that uh Tesla and
Whimo tend to be ahead of the rest and uh because
[17:41] they are the ones who put the most miles while
after the their the system was set up that way.
[17:47] Yeah. So this leads us to our last principle which
would be uh we see AI agents as a operational or
[17:56] organizational capability. So the question is why
is AI as a tool just not the the end goal here
[18:06] and why is that just not the solution for every
customer? Why is AI not the solution for every
[18:11] customer? Um well I would say you know seeing it
as calling an AI initiative probably not the best
[18:18] way to to to go about it. It's the best way to go
about it is to actually figure out what problem
[18:24] needs to be solved. And often times um the problem
that needs to be solved do not require LLMs. It
[18:32] um and sometimes it does and it really depends
upon you often time and and even if you do use
[18:39] LMS you don't always use them as the reasoning
portion of the system. Sometimes they are just
[18:45] a translation layer between systems and humans
there because the truth is is they don't reason.
[18:52] They don't reason. That's not what they're
doing. So you don't so you don't treat them
[18:57] like reasoning. You don't treat them as if they
are reason. They understand causality the same
[19:02] way that a human would understand causality. And
yeah, just went negative. So we'd love to continue
[19:08] this conversation. We have a lot to say about SPO.
We're we're a little over time now, but come visit
[19:14] our booth at 8:31. Booth 831. Uh we have plenty of
people there that can continue this conversation.
[19:21] Uh we'd love to hear what what you're
working on within this space. Thank you all.
