AI-Native Infrastructure

Intelligence
for the
Physical World

Physical AI 正在成為下一波 AI 產業的主戰場。AI 不再只存在於雲端——它正進入機器人、無人載具、EV、工廠與城市基礎建設。我們要打造的,是一套能讓 AI 進入真實世界的作業系統。

Explore Platform →
Book a Demo
NVIDIAAUTODESKAWSSIEMENSMicrosoft
Three Pillars, One Intelligence Layer
Pillar 01
Physical AI
VLA + WMA powered intelligence for robots, drones, EVs and more. Real-world reasoning and autonomy.
Pillar 02
Spatial Intelligence
AI-native 3D spatial workspace. Transform how industries design, build and operate.
Pillar 03
Business AI
AI that integrates data, systems and people for self-improving business operations.
RoboticsDronesAutonomous VehiclesManufacturingConstructionReal EstateEnergyBusiness
Finetune + RL · Model Training

我們深入訓練企業專屬模型

從通用 foundation model 出發,把企業資料、行業 know-how 與決策邏輯結構化進模型,再用 RL post-training loop 讓模型隨真實場景持續進化——這是 Lens AI 的護城河。

A
Phase 01 · Base
Enterprise Base Training
用企業真實語料把通用模型「拉進」企業語境
CPT
Continued Pre-Training
企業多年文件、SOP、紀錄、Code、報表
SFT
Supervised Fine-Tuning
專家標註的高品質指令/答案資料對
Tokenizer
行業專有詞彙、規範代號、產品料號擴詞
Output
會說企業語言的 Foundation 模型
B
Phase 02 · Vertical
Vertical Domain Encoding
把行業 know-how、角色邏輯、環境約束編進模型
Knowledge
Domain Knowledge Graph
圖紙、BIM、規範、料表、SOP 結構化
Role Logic
Multi-role Persona Embedding
建築師 / 機電 / 採購 / PM 的判準
Context
Dynamic Context Assembly
環境約束、合規規則、決策偏好
Output
能執行行業任務的 Domain Expert 模型
C
Phase 03 · RL Loop
RL Post-Training Loop
每一次任務都產生 reward signal,模型每天變強
Reward
Outcome-Aware Reward Model
業務 KPI、專家評分、規範通過率
RLHF
Expert RLHF + DPO
專家偏好直接優化 policy
Online
Continuous Online Learning
新案例 → reward → 每日 retrain
Output
越用越像企業專家的 Self-Improving 模型
The Moat
企業每一次採用 Lens AI,都會留下 reward dataexpert feedbackreal-world outcomes——這些資料只屬於這家企業的模型,競爭者無法複製
Compounding Advantage
Part 01

Physical AI

VLA + WMA — 從感知到行動,讓 AI 進入真實世界

The Hardware Wave · Why Now

Explosive growth — every system is a closed world

Billions pouring into robotics hardware — but every company ships a closed SDK, a proprietary stack, and zero interoperability.

Quadruped Robot Dog
$0.5B → $6.6B
2024 → 2032 · CAGR ~19–25%
Commercial UAV
$30B → $65B
2024 → 2032 · CAGR ~11–21%
Industrial Robot / Arm
$74B → $111B
2024 → 2030 · CAGR ~14%
Humanoid Robot
$1.6B → $15B
2024 → 2030 · CAGR ~17–40%
Unmanned Surface Vessel
$1.2B → $3.6B
2024 → 2034 · CAGR ~10–14%
AGV / AMR
$6.4B → $22B
2025 → 2030 · CAGR ~21–30%
$100B+ combined market · Every device speaks its own language · What connects them?
Market Gap · Why Now

Nobody is building this layer

Hardware is mature · Capital has flooded in · But one layer is still empty

Robot Dog
Commercial UAV
Industrial Arm
Humanoid
USV
AGV / AMR
APPLICATION LAYER, MISSING

No OS. No Protocol.
No Platform.

Every use case is built from scratch. The same problem solved a hundred times over.

Every Vertical
Rebuilt from zero
No Protocol
AI can't drive HW
No Platform
No dev ecosystem
Before Android, every mobile app had to handle Bluetooth, GPS, and camera drivers itself. Physical AI is at that exact inflection point right now.
THE ECOSYSTEM CONNECTOR

Lens OS for Physical AI

A unified agentic runtime. Any robot, drone, arm, or vessel speaks the same language. Build once, deploy on any hardware.

Quadruped
Unitree
Humanoid
Figure
UAV
DJI / Skydio
AGV / AMR
Geek+
USV
Saronic
Zero rip-and-replace
Sits above existing ROS2 / SDK. Zero changes to the underlying stack
One runtime, any vertical
Security, Agriculture, Logistics, Defense, Construction. One platform
All robots, one roof
Unified device integration that feels both effortless and complete
Solution · Lens OS Architecture

Lens Physical AI Architecture

從 edge device 到 GPU cloud,完整的 agent runtime — 三層架構讓每台機器人都能思考、抽象、行動

LAYER 01
01 / 03
Reasoning
Real-World Reasoning Layer
理解真實世界場景與環境。VLA 與 WMA 在這一層使用訓練模型讓 Robotic Device 理解自身環境——從感知到決策的完整推理鏈。
VLA WMA Fast LLM Strategy LLM
LAYER 02
02 / 03
Abstraction
Hardware Abstraction Layer
將控制系統與感測器轉換成 Agent 可以使用的各種工具。統一不同廠牌硬體的介面,讓 AI 專注於任務而非設備差異。
ROS2 Bridge SDK Wrapper Sensor API
LAYER 03
03 / 03
Action
CLI / Action Control Layer
整合不同工具、控制模組與工作流程——像用 LEGO 積木搭建系統。開發者用熟悉的 CLI 指令組裝任務,AI 自動編排執行。
Lens CLI Workflow Engine Plugin System
Philosophy
每一層都是為 AI 而設計——不是在舊系統上加 AI,而是 AI-native from ground up。Protocol 是 built on AI, built for AI。
Solution · Intelligence Engine

Three-Tier LLM Hierarchy

Reflex → Reasoning → Fleet Command — 三層 LLM 架構驅動即時反應到全域決策

Fast LLM
Fast LLM
The Reflex Agent
讀取 sensor context,在毫秒內推理並輸出 action code。體育場機器人看到可疑人物 → 立即觸發警報、鎖定目標、通知周邊設備。行為由 Strategy LLM 定義與管理。
Strategy LLM
Strategy LLM
The Reasoning Layer
每台機器人的大腦。負責建立、修改、刪除 Fast LLM — 決定監控什麼、如何反應。在 sandbox 模擬決策後才部署,串連歷史 log 的因果關係。
Supervisor LLM
Supervisor
Fleet Command
跨所有 Strategy LLM、所有機器人、所有 sensor 的全域視角。負責 fleet-wide 決策與安全閘門,將單機智慧轉化為集群協調。
Edge + Cloud Fast LLM 在 edge 即時推理,Strategy + Supervisor 在 cloud 深度思考。三層協同完成任務。
Demo · Physical AI

Physical AI in Action

Demo · Physical AI

Physical AI Demo

Example — Enterprise Security

Foxlink Group × Lens OS

Foxlink
Taiwan-listed · TWD 5.2B capital Dozens of subsidiaries across the group Robot dogs already deployed in internal security
Foxlink has deployed robot dogs in its internal security system. Lens OS acts as the unified agentic runtime — enabling the fleet to scale across factory floors, stadiums, and airports in parallel, without rewriting the underlying stack.
For Hardware Vendors
Multi-robot integration at minimal cost — all dogs managed via a unified protocol
For Developers
Easy AI + ROS integration — rules and AI chains work out of the box
Autonomous patrol routing — factory, stadium, airport
VLM real-time detection of suspects, weapons, and explosives
Multi-agent coordination — event-triggered reinforcement in seconds
New features added directly to Lens OS — fast iteration, no refactoring
Lens OS Physical Overview
Lens OS · Physical AI Runtime — Foxlink Robot Dog Fleet
Example — Construction & BIM Inspection

CAMTI Group × Lens OS + Lens CAD

CAMTI
Taiwan's #1 low-voltage systems integrator 30+ years in the industry 60+ subsidiaries 600+ buildings served
Lens OS connects UAV site surveying, robot dog inspection, and Lens CAD BIM comparison — forming a complete AI construction monitoring loop. Survey, compare, suggest — one runtime drives the entire workflow.
01
UAV Site Survey
UAV on-site flight — generates point cloud & live footage
02
Robot Dog BIM Inspection
Zone-by-zone patrol comparing progress against the BIM model
03
Lens OS CLI Loop
survey → compare → suggest
04
Lens CAD Live BIM Visualization
Deviation zones highlighted, construction reports auto-generated
Lens CAD · BIM Inspection
Insert Lens CAD demo video — UAV point cloud import, BIM deviation marking, robot dog patrol path
Lens CAD × Lens OS — CAMTI Construction Monitoring Solution
Example — Maritime Defense & Harbor Patrol

Carbon-Tek × Lens OS

Carbon-Tek
Taiwan defense supply chain · ~20 years established Focused on unmanned vehicle systems Primary applications in defense and aerospace
UAV + USV joint harbor patrol — aerial and surface vehicles executing missions together on a single agentic runtime. Defense-grade multi-vehicle autonomous patrol, already deployed in the defense supply chain.
UAV + USV joint harbor patrol coordination
24/7 perimeter anomaly detection and alerting
Cross-vehicle mission planning — single runtime orchestration
Defense-grade autonomy — live deployment in military scenarios
DEFENSE GRADE MARITIME
Carbon-Tek Unmanned Surface Vessel
Carbon-Tek USV — Unmanned Surface Vessel Harbor Patrol
Demo · Drone

Drone as an autonomous inspection agent

Demo · LEGO Build 1

Building intelligence from modular components

Demo · LEGO Build 2

Assembling the Physical AI stack

Part 02

Spatial Intelligence
Lens Builder

Lens Builder 是 Lens AI 在 Building / AEC 場景中的 Spatial Intelligence solution。

Spatial Intelligence · Pain Point

Building projects require every role to align on the same data foundation

一個建案同時牽涉業主、建築師、工程師、承包商、BIM 團隊與法務/財務角色

Current State
The Problem
每個角色都依賴不同資料:模型、圖面、法規、成本、物料、時程、現場紀錄
每個角色對同一個空間、同一個變更、同一個風險的判讀不一致
資料無法對齊,決策就會變慢
估價、送審、施工、變更、驗收都需要反覆溝通

The opportunity is not only automation.
It's building a shared intelligence layer.

Lens Builder
The Opportunity
統一空間資料基礎
多角色一致判讀
可追溯的決策鏈
AI-native 工程協作

Building the shared intelligence layer
for complex construction decisions.

Spatial Intelligence · Our Approach

Lens Builder creates a shared spatial intelligence layer

將 BIM、CAD、Floor Plan、需求書、法規、BOM、報價與排程對齊到同一個 Spatial Context。讓 AI 理解空間關係、元件參數、材料邏輯、成本結構與施工依賴。

Owner
成本、ROI、風險
Architect
法規、設計、送審
Engineer
衝突、管線、維修空間
Contractor
報價、工期、施工依賴
BIM Team
模型品質、座標、資訊完整度
Demo · BIM

BIM as a spatial decision graph

讀取 BIM 裡的設備、材料、管線、空間位置與物件參數
將模型中的元件關係轉成可分析的 spatial context
同一個 BIM 模型,可被轉換成不同角色的決策輸出
Demo 重點:不是展示模型,而是展示 AI 如何理解模型背後的工程意義
Takeaway

From BIM objects to role-specific decisions.

Demo · Floor Plan

Floor Plan as the entry point of spatial understanding

解析房間、牆面、門窗、動線與區域關係
將 2D floor plan 轉成 AI 可理解的空間結構
支援早期規劃、空間配置、裝修提案、需求比對與成本初估
Demo 重點:從低門檻圖面開始建立 building intelligence
Takeaway

Every building workflow starts from understanding space.

Spatial Intelligence · Vision

Building the spatial intelligence layer for the built environment

Part 03 · Business AI

Enterprises are ready for AI,
but not ready to operate with AI

企業正在快速導入 AI。老闆看得到效率提升的機會,也知道公司內部有大量流程可以被 AI 改造。但真正的問題是:企業不是缺一個 AI 工具,而是缺一套能讓 AI 進入營運流程的方法。

Lens AI 的 BusinessAI,是協助企業把內部營運與對客戶的服務流程,轉換成 AI-native / Agentic Flow。
Business AI · Pain Point

AI adoption creates new complexity before it creates efficiency

企業導入 AI 後,反而多了更多人、更多工具、更多流程要管理

The Reality
導入 AI 之後
新的人操作 AI 系統
新的人協助導入
新的 AI 主管接手成果
原本部門仍用舊流程協作
SaaS、表單、人工審核仍然存在

AI 沒有真正進入企業 operation。
它只是被放在既有流程旁邊,變成另一個工具。

Real Need
真正需要的是
重新整理流程
角色分工
任務邏輯
進一步 re-org

才能真正釋放 AI 效率。
這就是 Lens AI BusinessAI 的切入點。

Business AI · Our Role

Lens AI acts as the Forward Deployed Engineering team

我們進入企業現場,理解它的資料、流程、組織與商業目標

我們不相信通用型 BusinessAI 可以直接解決企業問題。每一家企業的流程、資料、組織與客戶關係都不同。

1
進入企業流程
2
找出高頻、低效、可被 AI 改造的環節
3
重新設計工作流
4
建立 AI 可以理解的 business context
5
將流程逐步轉換成 Agentic Flow
我們不相信通用型 BusinessAI 可以直接解決企業問題。每一家企業的流程、資料、組織與客戶關係都不同。
Solution 01

Business Knowledge Model & Runtime

Business Knowledge Model 會把企業內部的資訊轉換成 AI 可理解、可推理、可執行的 context。Runtime 則讓 AI agent 不只是回答問題,而是能在企業流程中執行任務。

把過去分散在 SaaS、Excel、Email、Slack、LINE、人工判斷裡的流程,逐步轉換成可被 AI 協作與自動化的系統。

商品與服務資料
客戶與交易紀錄
銷售、庫存、客服流程
SOP 與部門分工
任務、會議、Email、內部溝通紀錄
把過去分散在 SaaS、Excel、Email、Slack、LINE、人工判斷裡的流程,逐步轉換成可被 AI 協作與自動化的系統。
Solution 02

Generative BusinessAI & Lens OS

Lens OS 讓企業快速導入 AI,把內部營運流程或對客戶的服務流程,整合成 Agentic Flow。

不只是一次性導入專案,而是讓企業 operation 持續變得更 AI-native 的系統。

會議內容
Email
任務紀錄
客戶互動
業務流程
CEO 目標
公司策略
部門優先順序
不只是一次性導入專案,而是讓企業 operation 持續變得更 AI-native 的系統。
Case Study 01 · Education

陳立數學:企業教材解決方案

將教材、題庫、解題邏輯結構化
建立課程知識模型
協助老師快速備課
生成不同程度學生適用的教材
整合教學、測驗、複習與成效追蹤流程
讓教育企業的內容資產可以被 AI 持續理解、重組與生成
[ Screenshot / Demo — awaiting content ]
Case Study 02 · E-Commerce

網路書商:AI Marketing Flow

理解書籍內容與分類
分析讀者族群與購買行為
規劃活動與檔期
生成 Email、社群、廣告文案
根據庫存與銷售狀況調整推廣策略
追蹤成效並持續優化下一輪行銷
從人工企劃與單點工具,轉變成可持續運作的 Agentic Flow
[ Screenshot / Demo — awaiting content ]
Demo · BookStore

Sample 2 — BookStore

Business AI · Summary

Lens AI is building the operating layer for Agentic Enterprise

企業 AI 的下一階段,不只是更多工具,而是新的 operation layer。Lens AI 透過 FDE 進入企業現場,建立 Business Knowledge Model,部署 Runtime,並透過 Lens OS 將企業流程轉換成 Agentic Flow。

我們的目標是成為企業導入 AI 的核心基礎層。

Build the OS for the
Physical World

不只是做一台機器人,或是一台車。
而是打造一套能讓 AI 進入真實世界的作業系統。

Somer · CTO, Lens AI

Explore Platform
Book a Demo
NVIDIAAUTODESKAWSSIEMENSMicrosoft