From Digitalization toPhysical AI
从数字化到物理AI
Item's AI Journey Over the Past Year — Practices, Challenges & Cognitive Restructuring
Item过去一年的AI实践、挑战与认知重构
Item serves as the implementation foundation for all supply chain transformation projects across UNIS Group
Item 是 UNIS 集团所有供应链转型项目的实施底座
Who We Are
关于我们
UNIS provides omni-channel fulfillment; Item builds the technology platform that powers it all.
UNIS提供全渠道履约服务；Item构建驱动一切的技术平台。
www.unisco.com
Omni-channel fulfillment on-time and in-full. UNIS started serving Lenovo since 2019 at Memphis, US. Strategically designed national footprint reaches 98% of consumers with same-or next-day service.
全渠道准时足量履约。UNIS自2019年起在孟菲斯为联想提供服务，全国性网络覆盖98%消费者的当日或次日达。
100M+
Orders / Year
年订单量
1,200+
Customer Accounts
客户账户
$20B
Inventory Value
库存价值
10M+
Sq Ft Warehouse
平方英尺仓库
The technology engine behind UNIS. Item is the implementation foundation for all supply chain transformation projects across UNIS Group.
UNIS背后的技术引擎。Item是UNIS集团所有供应链转型项目的实施底座。
CAPABILITY STACK
SaaS Platform
OMS · WMS · TMS · YMS · RMS · Billing
AI Agent Factory
Ontology → Agent Assembly → Runtime
Robotics Integration
WES · WCS · AMR · AS/RS · Sortation
"From digital operations → AI intelligence → physical automation"
"从数字化运营 → AI智能化 → 物理自动化"
CUSTOMERS & PARTNERS
AI Evolution Path & Scale Validation
AI 演进路径与规模验证
A three-phase evolution framework from digitalization to Physical AI, backed by production-validated metrics.
从数字化到物理AI的三阶段演进框架，以及经过生产验证的规模数据。
Digitalization
数字化
Full-stack SaaS (OMS/WMS/TMS/YMS) + Digital Twin, making business processes machine-readable.
全栈SaaS（OMS/WMS/TMS/YMS）+ 数字孪生，让业务变得可被系统理解。
AI Agent
AI Agent
Agent orchestration + contextual memory, enabling decisions to be automatically executed.
智能体编排 + 上下文记忆，让决策变得可被自动执行。
Physical AI
物理 AI
WES/WCS + robotics (AMR, AS/RS, sorters) deep integration — intelligence leaves the screen.
WES/WCS + 各类机器人（AMR、AS/RS、分拣系统）深度整合，智能离开屏幕。
Annual Orders Processed
年处理订单量
Legacy WMS Instances Unified
统一的传统WMS实例
Warehouse Coverage
仓库覆盖面积
Multi-Agent Collaboration in Action
多Agent协作实录
A breakthrough achieved in recent days — multiple AI Agents working in concert, truly bridging the digital and physical worlds.
最近几天的突破性成果——多个AI Agent协同工作，真正打通了数字世界与物理世界的边界。
Agent Orchestration
Agent 协同编排
Specialized Agents each handle their domain — from task planning and path optimization to real-time execution, forming a complete decision-execution loop.
多个专业Agent各司其职——从任务规划、路径优化到实时执行，形成完整的决策-执行闭环。
Digital ↔ Physical Bridge
数字↔物理打通
AI Agent decisions no longer stay on screen — they directly drive robots and devices in the physical world to complete real operations.
AI Agent的决策不再停留在屏幕上，而是直接驱动物理世界的机器人和设备完成真实作业。
Real-time Feedback Loop
实时反馈回路
Sensor data from the physical world feeds back in real-time, enabling Agents to dynamically adjust strategies for true closed-loop intelligence.
物理世界的传感器数据实时回传，Agent动态调整策略，实现真正的闭环智能。
Office Automation — Agent-ification
Office Automation Agent化
From 'system of record' to 'system of action' — our experiments and results in Agent-ifying internal office and operational workflows.
从「系统记录」到「系统行动」的转变——我们在企业内部办公和运营流程中的Agent化尝试与成果。
Call Center Ticketing
In-house AI Customer Service System
自研AI客服系统
AI Agents understand customer intent and directly invoke OMS/WMS APIs for status queries and exception handling, significantly reducing manual intervention.
AI Agent不仅能理解客户意图，还能直接调用底层OMS/WMS接口进行状态查询和异常处理，大幅降低人工干预率。
Personal Assistant
Evolution from In-house to OpenClaw
从自研到OpenClaw的演进
Initially built a general-purpose personal assistant Agent, later replaced by OpenClaw. Delegated generic tasks to specialized tools, refocusing internal efforts on domain depth.
早期尝试构建通用型个人助理Agent，后替换为OpenClaw。将通用任务交由专业工具处理，内部精力聚焦业务深度。
Driver AI Agent
Full-process Agent for Drivers
司机端全流程Agent化
Equips drivers with AI assistance in TMS/YMS workflows — handling check-ins, dock congestion, missing documents, and providing real-time guidance.
在TMS/YMS环节为司机配备AI辅助系统，处理日常签到、月台拥堵、文件缺失等异常情况，提供实时指导。
Challenges & Lessons Learned
问题与心路历程
An honest account of pitfalls and cognitive shifts — the awakening journey from blind faith in LLMs to architectural restructuring.
坦诚分享踩过的坑和认知转变——从迷信大模型到重构架构的觉醒之路。
早期陷阱
Over-reliance on LLM "General Capabilities"
迷信大模型的「通用能力」
We attempted to solve all complex business problems purely through Prompt Engineering. This led to severe hallucination issues, unreliable execution, and a complete inability to handle enterprise compliance red lines and SOP requirements.
我们曾试图单纯通过Prompt Engineering来解决所有复杂的业务问题。这种做法导致了严重的幻觉问题，系统执行不可靠，且完全无法应对复杂的企业合规红线和SOP要求。
认知转折
Foundation Models Are Not the Moat
基础模型不是护城河
We realized that as models rapidly iterate and go open-source, general capabilities quickly commoditize. The true moat lies in Business Harness and Ontology.
我们意识到，随着模型的快速迭代和开源，通用能力会迅速商品化甚至贬值。真正的护城河在于业务Harness和本体知识库（Ontology）。
Architecture Restructuring Direction
架构重构方向
Fully decouple the brain (LLM) from the hands (execution tools), pivoting entirely toward building robust Business Harness and Ontology knowledge bases.
把大脑（大模型）与手脚（执行工具）彻底解耦，全面转向构建强大的「业务Harness」和「Ontology知识库」。
Ontology: The Semantic Backbone of Enterprise AI
Ontology：企业AI的语义骨干
General-purpose LLMs don't understand enterprise 'domain dark knowledge' or SOPs. Without Ontology, AI Agents cannot achieve accuracy and behavioral consistency in enterprise scenarios.
通用大模型不懂企业的「行业暗知识」和「SOP」。没有Ontology，AI Agent无法在企业级场景中做到准确和行为一致。
Layer 2 · Neo4j GraphRAG · Enterprise Brain 企业大脑
What is Ontology?
什么是Ontology？
Ontology is more than a knowledge graph. It is a domain-specific structural framework, entity model, and semantic backbone. It defines relationships, constraints, and reasoning rules among all concepts within a domain, enabling AI Agents to think and act like domain experts.
Ontology不仅仅是知识图谱。它是特定领域的结构化框架、实体模型和语义骨干。它定义了领域内所有概念之间的关系、约束和推理规则，让AI Agent能够像领域专家一样思考和行动。
Construction Results
构建成果
Build Time
构建时间
Research Iterations
自主研究迭代
Domain Coverage
领域知识覆盖
Why is Ontology Decisive?
为什么Ontology是决定性的？
- Unifying 55 independent WMS instances into a single semantic architecture
将55个独立WMS实例统一到一个语义架构中
- Enabling AI Agents to understand the real meaning of domain concepts like SKU, Dock, and Wave
让AI Agent理解「SKU」「月台」「波次」等领域概念的真实含义
- Providing validation benchmarks for Harness — Agent outputs must conform to Ontology-defined constraints
为Harness提供验证基准——Agent的输出必须符合Ontology定义的约束
- The secret weapon for transforming Agents from 'toys' into 'productivity tools'
将Agent从「玩具」转化为「生产力工具」的秘密武器
"Ontology is the bridge from general intelligence to domain expertise. Without it, LLMs are merely clever but ignorant outsiders."
"Ontology是让AI从通用智能走向专业智能的桥梁。没有它，大模型只是一个聪明但无知的外来者。"
Harness: Constraints & Moat for AI
Harness：AI的约束与护城河
Agent = Model + Harness. Harness is the structured software environment surrounding the model, designed to guide, constrain, and validate AI outputs.
Agent = Model + Harness。Harness是包围在模型外围，用于引导、约束和验证AI输出的结构化软件环境。
Generic Harness
通用 Harness
Infrastructure-level capabilities: sandbox isolation, credential management, crash recovery, context management, MCP protocol, etc.
沙盒隔离、凭证管理、崩溃恢复、上下文管理、MCP协议等基础设施级能力。
Depreciates with model upgrades — belongs to model vendors' territory, gradually becoming commoditized infrastructure
随模型升级而贬值 — 属于模型原厂领地，像水电煤一样会逐渐基础设施化
Business Harness
业务 Harness
Domain dark knowledge, compliance red lines, vertical SOPs, organizational taste, and experience distillation flywheel.
行业暗知识、合规红线、垂直领域SOP、组织品味、经验蒸馏飞轮。
Appreciates with model upgrades — the exclusive territory of enterprises and practitioners, forming the true moat
随模型升级而增值 — 企业和从业者的专属领地，构成真正的护城河
Item's strategic focus is on building deep Business Harness, ensuring platform value won't be displaced by stronger LLMs, but instead compounds as underlying models improve.
Item的战略重心在于构建深厚的「业务Harness」，确保平台价值不会被更强的大模型取代，反而因底层模型变强而产生复利增值。
LLM Selection Strategy
大模型选型策略
A pragmatic Hybrid Router strategy, not locked to any single model — highly aligned with Lenovo's 'Hybrid AI' philosophy.
务实的混合路由策略（Hybrid Router），不绑定单一模型——与联想「Hybrid AI」理念高度一致。
US Frontier Models
High-complexity Reasoning / Code Gen
美国头部大模型 — 高复杂度推理 / 代码生成
复杂推理 / Agentic Workflow
长文本分析 / 代码生成
多模态理解 / 视频分析
China Open-source / Local Models
Privacy / Localization / Cost
中国开源/本土模型 — 隐私 / 本地化 / 成本敏感
本地化部署 / 中文理解
成本敏感场景 / 推理
垂直领域专用
Future Expansion: From LLM to Multimodal
未来扩展方向：从LLM到多模态
LMMs
Large Multimodal Models
大多模态模型
Unified understanding of text + image + audio
文本+图像+音频的统一理解
LVMs
Large Vision Models
大视觉模型
Visual monitoring / Video QC / Packing station QC
视觉监控 / 视频质检 / 打包台QC
LPMs
Large Physical Models
大物理模型
Robot control / Physical world interaction
机器人控制 / 物理世界交互
Hardware & Computing Infrastructure
硬件与算力基础设施
Evolution from cloud to edge — supporting massive Agent clusters and physical execution requires more than software alone.
从云端到边缘的演进——支撑庞大的Agent集群和物理执行，仅靠软件是不够的。
Current Hardware Foundation
当前硬件基础
Cloud GPU
云端GPU
LLM inference (OpenAI / Claude API)
大模型推理（OpenAI / Claude API）
Kafka + K8s
Kafka + K8s
Cloud-native stream processing & container orchestration
云原生流处理与容器编排
Edge Devices
边缘设备
AI camera data processing
AI摄像头数据处理
Edge Inference Chips
边缘推理芯片
NPU / LPU
Mass deployment of AMRs, shuttle systems, and autonomous forklifts requires edge chips designed for low-latency inference.
AMR、四向穿梭车、无人叉车的大规模部署需要专为低延迟推理设计的边缘芯片。
Cloud GPU Clusters
云端GPU集群
A100 / H100 / Domestic Alternatives
Supporting Agent Factory and the growing Ontology graph computation (GraphRAG).
支撑Agent Factory和不断增长的Ontology图谱计算（GraphRAG）。
High-speed Networking
高速网络硬件
5.5G / Wi-Fi 8 / UWB / Fiber
Real-time control for Physical AI demands ultra-low latency, requiring integration with next-gen networking hardware.
物理AI的实时控制对延迟要求极高，需要扩展对新一代网络硬件的集成支持。
Entering the Physical World
走向物理世界
Hardware-agnostic WES (Warehouse Execution System) — not locked to any single hardware vendor, bringing intelligence beyond the screen.
硬件无关的WES（仓库执行系统）——不绑定任何单一硬件厂商，让智能真正离开屏幕。
HARDWARE-AGNOSTIC WES
Software stack running stably in real warehouse production environments covering 10M+ sq ft
软件栈在超过1000万平方英尺的真实仓库生产环境中稳定运行
Integrated Robotic Equipment
已集成的机器人设备
AirRob AMR
AirRob AMR
Autonomous Mobile Robot
自主移动机器人
4-way Shuttle
四向穿梭车
High-density Storage
高密度存储
Bluecore AGV
Bluecore AGV
Automated Guided Vehicle
自动导引车
Autonomous Forklift
无人叉车
Automated Transport
自动搬运
Vision & Edge AI Applications
视觉与边缘AI应用
Staff Monitoring Agent
员工监控Agent
Behavior analysis & safety compliance
行为分析与安全合规
AI Camera Audit
AI摄像头审计
Packing station video QC
打包台视频QC
YMS Gate Recognition
YMS道闸识别
Intelligent vehicle recognition & scheduling
车辆智能识别与调度
5-LAYER AI ARCHITECTURE · 五层AI架构
Collaboration Vision
合作展望
Joint value with Lenovo — a strategic path from internal validation to external go-to-market.
与联想的联合价值——从内部验证到对外输出的战略路径。
Ecosystem Integration
生态融合
Deep ecosystem integration of Item's software stack with Lenovo's edge computing hardware and robotics, co-building 'software-defined hardware' solutions.
将Item的软件栈与联想的边缘计算硬件及机器人进行深度生态融合，共建「软件定义硬件」的解决方案。
Joint Reference Cases
联合参考案例
Starting with Lenovo's internal AMR Zone Picking project, jointly refining solutions and validating business value.
以联想内部的AMR Zone Picking项目作为起点，共同打磨联合解决方案并验证商业价值。
External Go-to-Market
对外输出
Beyond serving Lenovo's own supply chain upgrades, exporting joint solutions externally to co-build next-gen smart warehouse infrastructure.
不仅服务于联想自身的供应链升级，更要将联合解决方案向外输出，共同构建下一代智能仓库基础设施。
As the implementation foundation for all supply chain transformation projects across UNIS Group, Item has validated the complete path from digitalization to Physical AI in production environments. We look forward to partnering with Lenovo, deeply integrating this battle-tested AI middleware layer with Lenovo's hardware ecosystem, and jointly defining the standard for next-generation smart warehouses.
Item 作为 UNIS 集团所有供应链转型项目的实施底座，已经在生产环境中验证了从数字化到物理AI的完整路径。我们期待与联想携手，将这套经过实战检验的AI中间件层与联想的硬件生态深度融合，共同定义下一代智能仓库的标准。
Item AI Strategy Briefing — Confidential — UNIS Group