title: "AI-Native Vehicle Cabin:大螢幕、投影、Spatial Audio、語音、Agent 與環境控制" subtitle: "AI-Native Vehicle Cabin: A Multimodal Runtime Architecture for Displays, Projection, Spatial Audio, Agents, and Environmental Control" author: "Neo.K(EVEMISS / EveMissLab)" ai_collaboration: "Aletheia(GPT-5.6 Sol)" version: "0.1" status: "Research Paper / Canonical Source" date: "2026-08-24" language: "zh-TW" series: "Series B|Programmable Mobility & Spatial Continuum" series_number: "B-05" canonical_source: true encoding: "UTF-8"
AI-Native Vehicle Cabin:大螢幕、投影、Spatial Audio、語音、Agent 與環境控制
AI-Native Vehicle Cabin: A Multimodal Runtime Architecture for Displays, Projection, Spatial Audio, Agents, and Environmental Control
系列: Series B|Programmable Mobility & Spatial Continuum
編號: B-05
作者: Neo.K(EVEMISS / EveMissLab)
AI 協作: Aletheia(GPT-5.6 Sol)
版本: v0.1
日期: 2026-08-24
文件狀態: Research Paper / Canonical Source
摘要
B-01 將交通工具重新定義為 Programmable Mobility Space;B-02 將交通時間拆解為可回收與可配置的活動時間;B-03 定義 Mobility Cognitive Continuity 與 Mobility Resume Capsule;B-04 則建立 Spatial Attention Zones,要求所有車內互動先通過角色、安全、注意力、隱私與權限約束。本篇將前四篇真正組裝成一個可實作的車艙執行層:
AI-Native Vehicle Cabin(ANVC)
AI 原生多模態車艙
本文主張,「AI-native cabin」不是在傳統 infotainment 上多加一個聊天機器人,也不是把所有控制都交給一個 LLM。它是一套把:
- display;
- projection;
- audio;
- microphone;
- speech;
- camera / occupant sensing;
- seat;
- lighting;
- climate;
- privacy devices;
- communication continuity;
- AI agents;
- network / cloud;
- vehicle-safe cabin controls;
統合成可被意圖、活動狀態與 Zone Contract 編排的 runtime。
本文定義:
其中:
- :display / visual surfaces;
- :projection / augmented visual output;
- :audio / spatial sound zones;
- :microphone / speech input;
- :environment actuators;
- :AI / agent runtime;
- :communication / continuity state;
- :Spatial Attention Zone contracts;
- :privacy / authority;
- :safety / vehicle-state boundary。
本文第一個核心命題是:
2026 年的產業工程已清楚呈現「車艙 AI」與「駕駛/自動駕駛」可以共存但必須分層:NVIDIA 的 in-vehicle AI agent architecture 明確包含 ASR、orchestrator、LLM/VLM、tool use、TTS、edge / cloud pipeline 與 policy enforcement;Qualcomm Snapdragon Cockpit 平台則提供 multimodal LLM、多高解析度顯示、zonal audio、active noise / echo cancellation、driver monitoring 與 edge AI;Mercedes 2026 年也推進 on-device speech、language understanding 與 reasoning。這些能力證明「車艙本身成為 AI compute / multimodal runtime」已是現代工程方向,但不代表 cabin agent 應直接取得 safety-critical driving control。
本文第二個核心命題是:
使用者不需要指定:
把第 3 號燈調到 18%,把右後座音響 gain 調多少。
他可以表達:
我要休息。
或:
我要繼續工作。
Experience Compiler 將高階意圖轉成 Cabin Experience Plan,再由 B-04 的 Zone Contract、安全政策、能力註冊表與 actuator contract 決定哪些控制允許執行。
第三個核心命題是:
成熟車艙不應同時把所有 display、speaker、mic、camera、AI notification 全開,而是依 activity、seat、privacy、motion、attention、network 與 safety 選擇最小充分模態。若人在 driver zone,dense PDF 可能被轉成簡短語音或 defer;若在 rear passenger zone,則可恢復完整大螢幕、鍵盤與私有音訊。
第四個核心命題是:
當系統可以控制 lighting、seat posture、temperature、airflow、privacy glass、audio field、display、projection 與 notification 時,UI 已不再只存在於螢幕。車艙環境本身成為一個可程式輸出域。這使 Work、Rest、Entertainment、Social、Creative、Idle 等 activity mode 可以被編譯成不同的 physical / digital state。
本文提出 Experience Intent、Cabin Capability Registry、Cabin Experience Plan、Cabin Control Contract、Actuation Receipt、Zone Binding、Device Driver Boundary、Edge/Cloud AI Split、Graceful Degradation、Shared-Vehicle Teardown 等核心規格,並給出 Work、Rest、Entertainment、Social 與 Agent-Only 五個參考 profile。
本文最終將車艙 runtime 收斂為:
B-06 將在這個技術執行層之上研究 Mobility Hospitality 與 Robotaxi Economy:當車艙不再只是「有沒有冷氣、有沒有司機」,而是具有不同可程式時空品質時,移動服務如何形成新的產品分級、定價與旅程價值。
Abstract
An AI-native vehicle cabin is not merely an infotainment system with a conversational model attached. It is a multimodal runtime that coordinates displays, projection, audio zones, microphones, speech, environmental controls, AI agents, communication continuity, privacy, and vehicle-safe actuators under explicit spatial-attention and safety constraints.
This paper introduces the AI-Native Vehicle Cabin (ANVC) architecture. It treats the cabin as a set of programmable output and input surfaces governed by an Experience Compiler. High-level human intent such as WORK, REST, ENTERTAINMENT, SOCIAL, or IDLE is compiled into a zone-constrained Cabin Experience Plan. The plan is then translated through bounded device contracts rather than exposing raw vehicle control interfaces directly to an AI model.
The architecture explicitly separates cabin intelligence from driving authority. Current 2026 industry systems already demonstrate many required primitives: agentic multimodal AI pipelines, edge-cloud orchestration, multiple high-resolution displays, zonal audio, active noise control, conversational AI, contextual personalization, and on-device speech / language reasoning. These are treated as engineering anchors, not as proof that a complete programmable mobility space already exists.
The paper defines a Cabin Capability Registry, Experience Intent, Cabin Experience Plan, Cabin Control Contract, Actuation Receipt, edge/cloud execution split, graceful degradation, and shared-vehicle teardown. It also establishes that the physical environment itself is a first-class output surface: lighting, seat posture, climate, airflow, privacy glass, audio fields, and visual surfaces may all participate in a cabin mode.
The resulting runtime composes B-03's Mobility Resume Capsule and B-04's Zone Contract into an executable cabin state. Series B-06 then uses this technical foundation to model Mobility Hospitality and robotaxi service differentiation.
0. 車艙不是一個 App
傳統思維:
本文改寫:
Infotainment 只是其中一個子系統。
1. AI-Native 不等於裝一個 Chatbot
如果:
Existing IVI
+ Chat Window
那只是:
真正 AI-native cabin 必須讓 AI Runtime 能理解:
- occupant;
- activity;
- zone;
- privacy;
- network;
- cabin capability;
- safety;
- continuity state。
2. 車艙狀態
定義:
3. — Display Surfaces
包括:
- center display;
- passenger display;
- rear display;
- foldable display;
- seatback display;
- ceiling display;
- personal display;
- wearable display。
4. — Projection / Augmented Visual Surface
包括:
- HUD;
- windshield projection;
- panoramic projection;
- cabin projection;
- AR overlay。
BMW Panoramic iDrive 等現代系統已經展示:
不再只是傳統儀表。
5. — Audio / Spatial Sound
包括:
- multi-channel audio;
- zonal audio;
- headrest speakers;
- spatial audio;
- ANC;
- echo cancellation;
- masking;
- private audio。
Qualcomm Snapdragon Cockpit 目前直接列出:
- premium sound;
- zonal audio;
- active noise / echo cancellation。
6. — Microphone / Speech
包括:
- cabin microphone;
- seat microphone;
- beamformed microphone;
- wake-word;
- VAD;
- ASR;
- barge-in;
- speech diarization;
- privacy capture policy。
7. — Environment Actuators
包括:
- lighting;
- seat position;
- seat heating / cooling;
- climate;
- airflow;
- window;
- privacy glass;
- ambient sound;
- display brightness;
- physical partition if available。
8. — AI / Agent Runtime
包括:
- local model;
- cloud model;
- ASR;
- TTS;
- VLM;
- planner;
- tool router;
- personal agent;
- cabin agent;
- continuity guardian。
9. — Communication / Continuity
承接 Series A:
- identity;
- session;
- task;
- memory;
- artifact;
- provider;
- network;
- authority。
10. — Spatial Attention Zones
承接 B-04。
Zone 決定:
- 哪個 Surface;
- 哪個 modality;
- 哪種 complexity;
- 哪個 authority;
此刻可用。
11. — Privacy / Authority
包括:
- audience;
- data class;
- local retention;
- personal account;
- cabin sharing;
- control scope;
- teardown。
12. — Safety Boundary
包括:
- driving state;
- occupant role;
- motion;
- emergency;
- OEM safety policy;
- certified restrictions。
13. Cabin Intelligence ≠ Driving Authority
本篇最重要:
即使 vehicle compute 越來越 centralized,
也不代表 software authority 必須合併。
14. Central Compute ≠ Central Authority
2026 Qualcomm 已展示:
- cockpit;
- driver assistance;
- body;
- connectivity;
跨 domain central compute。
但:
這是 architecture safety principle。
15. Why This Matters
如果 cabin agent:
幫我快一點。
不能直接:
它最多形成:
再交由合法 driving / navigation system 處理。
16. Experience Compiler
定義:
17. Experience Intent
例如:
intent: "REST"
duration_min: 40
privacy: "private"
allow_interruptions:
emergency: true
work: false
preferred:
light: "low"
audio: "quiet"
seat: "reclined"
18. Work Intent
intent: "WORK"
activity_ref: "paper_17"
resume_capsule: "mrc_123"
preferred:
display: "large"
audio: "private"
input:
- "voice"
- "keyboard"
network_priority: "interactive"
19. Experience Plan
Compiler output:
plan_id: "xp_..."
zone_id: "rear_right"
activity_mode: "WORK"
visual:
target_display: "rear_private_main"
layout: "document_focus"
audio:
zone: "rear_right_headrest"
mode: "private"
input:
voice: true
keyboard: true
environment:
light: "work_neutral"
seat: "upright"
climate: "personal_profile"
notification:
work: "normal"
personal: "summary"
20. Plan 不是 Raw Command
Experience Plan:
而不是:
GPIO 4 = HIGH
CAN 0x123 = ...
21. Cabin Control Contract
Plan 之下:
每個 actuator 必須宣告:
- capability;
- range;
- safety class;
- owner;
- authority;
- rollback;
- evidence。
22. Capability Registry
例如:
capability_id: "seat.recline"
zone_scope:
- "rear_right"
range:
min: 0
max: 35
safety_class: "CABIN_COMFORT"
requires:
occupant_detected: true
seatbelt_policy: "compatible"
23. Display Capability
capability_id: "display.rear_private_main"
resolution: "4K"
input_modes:
- "touch"
- "keyboard"
privacy_class: "PRIVATE"
driver_exposure: "LOW"
24. Audio Capability
capability_id: "audio.headrest_rr"
zone: "rear_right"
features:
- "zonal_playback"
- "anc"
- "echo_cancel"
leakage_class: "MEDIUM"
25. AI Capability
capability_id: "ai.cabin_agent"
execution:
- "edge"
- "cloud"
inputs:
- "speech"
- "context"
- "telemetry"
allowed_actions:
- "cabin_comfort"
- "media"
- "communication_intent"
forbidden_actions:
- "raw_driving_control"
26. Industry Primitive:Multimodal LLM
Qualcomm Cockpit 目前直接把:
列為 cockpit capability。
其 personalization 可以結合:
- user;
- context;
- voice;
- touch;
- audio。
27. Industry Primitive:Multiple High-Res Displays
Snapdragon Cockpit 平台列出:
- front displays;
- rear displays;
- high-resolution graphics。
2026 Leapmotor / Qualcomm centralized solution 甚至支援:
28. Industry Primitive:18-Channel Audio
同一 2026 platform 宣稱:
作為 immersive in-car entertainment 的能力。
這不等於 personal sound zone 完成,
但硬體自由度已經很高。
29. Industry Primitive:Agentic AI
NVIDIA 2026 明確描述:
包含:
- reasoning;
- planning;
- acting;
- memory;
- multimodal;
- proactive;
- personalized。
30. NVIDIA Agent Pipeline
其代表性 pipeline:
並可:
31. Edge AI Box 的意義
NVIDIA 2026 還提出:
既有 IVI 可透過獨立 AI Box 升級 agentic AI。
這對 PMS 很重要。
因為:
不一定要求整台車從零重新設計。
32. 這與我們早先「舊車也可以」一致
弱版本:
就可以建立部分:
33. 新車則可以 Zonal / Central Architecture
新一代 SDV:
- central compute;
- zonal controllers;
- OTA;
- service-oriented architecture。
更適合:
34. Mercedes On-Device AI
2026 Mercedes + Liquid AI:
- on-device speech;
- language understanding;
- reasoning;
- privacy;
- reduced cloud dependence。
所以:
35. Edge / Cloud Split
定義:
36. Edge Responsibilities
優先:
- wake word;
- basic ASR;
- privacy-sensitive inference;
- cabin controls;
- low-latency command;
- offline fallback;
- safety-related gating。
37. Cloud Responsibilities
適合:
- deep research;
- large-model reasoning;
- external search;
- enterprise access;
- long-context analysis;
- cross-device continuity。
38. Cloud Failure
若:
車艙不能:
應降階。
39. Graceful Degradation
例如:
40. Modality Degradation
網路差:
但:
不能降階。
41. Multimodal ≠ Everything On
成熟:
不是:
42. Modality Selection
由:
決定。
43. Example:Driver
活動:
Zone:
Decision:
44. Example:Rear Passenger
同一 activity:
Zone:
Decision:
45. Environment Is First-Class Output
本文核心:
46. Work Profile
可能:
- seat upright;
- neutral light;
- private display;
- focused audio;
- high network priority;
- reduced entertainment notifications。
47. Rest Profile
- reclined seat;
- low light;
- display off;
- DND;
- low audio;
- emergency-only interrupt。
48. Entertainment Profile
- large display;
- spatial audio;
- ambient light;
- game input;
- lower work notification。
49. Social Profile
- shared display;
- shared audio;
- conversation priority;
- AI low-interruption。
50. Creative Profile
- drawing / writing surface;
- voice capture;
- camera if allowed;
- AI co-creation;
- artifact persistence。
51. Idle Profile
最重要之一:
Do nothing.
System:
- 不推薦工作;
- 不拉起任務;
- 不主動最佳化;
- 只保留安全 / emergency。
52. Agent-Only Profile
B-02:
車艙可以:
- human rests;
- agent researches;
- no work UI;
- arrival review packet。
53. Environment Compiler
可分:
Digital:
- app;
- display;
- media;
- communication。
Physical:
- light;
- seat;
- climate;
- audio field。
54. Physical Planner Must Be Bounded
所有 physical output:
例如 seat:
不能超過 hardware / safety range。
55. Cabin Actuation
定義:
56. Raw Actuation Is Not Semantic Intent
所以:
需要 compiler / controller。
57. Actuation Receipt
每次重要狀態變更:
receipt_id: "cabin_rcpt_..."
plan_id: "xp_..."
zone_id: "rear_right"
capability_id: "seat.recline"
requested: 20
applied: 20
status: "SUCCEEDED"
safety_policy: "seat-v4"
58. Why Evidence Matters
AI 說:
已經調暗燈光。
不等於燈真的調暗。
所以:
59. Cabin State Feedback
控制後:
例如:
- seat sensor;
- climate state;
- display status。
60. Closed-Loop Cabin Control
不是:
61. Display Orchestration
多 display:
需要:
- ownership;
- privacy;
- zone;
- content type;
- driver exposure;
- priority。
62. Main Display ≠ Global Display
一個中心大螢幕:
可能只適合:
- navigation;
- shared control。
私人內容應:
63. Projection Orchestration
windshield / HUD:
不能拿來播放任意工作內容。
64. Passenger Projection
rear / side projection:
可更自由。
所以:
也必須有 Zone Contract。
65. Audio Orchestration
音訊可以是:
- shared;
- zone;
- headset;
- safety channel;
- assistant channel。
66. Audio Priority
可定義:
但 user policy 可微調非 safety 類。
67. Safety Audio Ducking
若:
則:
但私人敏感內容不應因此播到 shared speaker。
68. Microphone Orchestration
Mic:
- wake;
- communication;
- cabin command;
- safety sensing。
不同用途應不同 data policy。
69. Mic Purpose Binding
70. Camera / Vision
Cabin vision 可以支援:
- occupant detection;
- gesture;
- attention;
- safety;
- personalization。
但:
71. On-Device Privacy
Mercedes 2026 on-device AI 與 NVIDIA edge-first architecture 都支持:
作為合理方向。
72. Local Memory Cache
shared vehicle:
只能保存:
Ride end:
清除。
73. Shared Vehicle Teardown
正式:
74. Personal Car Teardown
私人車不同。
可以保留:
- profile;
- preferences;
- trusted device;
但仍應:
- encrypted;
- revocable;
- exportable。
75. Driver vs Passenger Account
同一車:
不能共用:
- mail;
- AI memory;
- history。
76. Multi-User AI
可以:
每個 occupant 有:
- own agent;
- own context;
- own permission。
77. Cabin Agent
另有:
它是 shared system agent。
78. Personal Agent vs Cabin Agent
Personal Agent:
- knows user;
- owns personal context;
- plans tasks。
Cabin Agent:
- knows vehicle capability;
- manages cabin environment;
- enforces zone-safe execution。
79. Agent Separation
因此:
80. Agent-to-Agent Intent
Personal Agent:
Neo.K 想休息 40 分鐘。
Cabin Agent:
- checks zone;
- checks capability;
- builds plan;
- actuates allowed environment。
81. Cabin Agent Does Not Need Personal History
它只需要:
而不是:
把整個個人 memory 丟給車。
82. Privacy by Contract Minimization
傳輸:
intent: "REST"
temperature_preference: 23
light: "low"
而不是:
full personal profile
full email history
full AI memory
83. AI Memory Boundary
車艙本地 AI 可以有:
Ride end:
除非明確保存。
84. Continuity Memory Remains External
正式 personal continuity state:
車只拿 temporary projection。
85. Source-of-Record Boundary
車上看到/聽到的:
- summary;
- visualization;
- TTS;
都只是:
Canonical artifact 在 Series A state domain。
86. Vehicle Surface as Ephemeral Renderer
shared robotaxi 可視為:
不是:
87. Surface Attestation
更高階可以要求:
- device attestation;
- software version;
- privacy mode;
- local erase capability。
再決定是否投影 D3 / D4 content。
88. Network Orchestration
車上:
- 5G;
- Wi-Fi;
- satellite;
- tethering;
都可以進 Series A AHCF。
Cabin Runtime 只要求:
89. Work Flow Requirement
例如:
class: "INTERACTIVE_WORK"
latency: "medium"
bandwidth: "high"
privacy: "high"
fallback:
- "audio"
- "offline_draft"
90. Entertainment Flow Requirement
class: "STREAMING"
bandwidth: "high"
latency: "medium"
fallback:
- "lower_resolution"
91. Network-Aware Experience
若 bandwidth 降:
4K:
但 activity identity 不變。
92. Compute-Aware Experience
edge compute limited:
cloud down:
93. Model-Aware Experience
不同 AI provider:
能力不同。
Experience Compiler 只應要求:
不鎖 provider。
94. Cabin Profile Portability
使用者 profile:
REST:
seat 20°
temp 23
light low
audio private
應該:
這會直接接 B-06 service economy。
95. Capability Matching
不同車:
Profile compiler:
96. Graceful Profile Degradation
如果車沒有:
- privacy glass;
則:
97. If Capability Insufficient
如果:
則:
98. Motion-Aware Cabin
急彎:
- visual reading 可能暫停;
- screen layout simplification;
- game pause。
99. Motion Input
Cabin Runtime 可以使用:
- acceleration;
- turn;
- road roughness;
- route phase。
但只做:
不取得 driving control。
100. Motion Sickness Adaptation
若偵測:
可:
- reduce visual motion;
- shrink immersive display;
- switch to audio;
- change seat orientation if allowed;
- pause activity。
101. Environment as Feedback
若乘員說:
太亮。
AI 不只 reply:
好的。
而是:
但仍需要:
- capability;
- authority;
- receipt。
102. Natural Language to Environment
這就是:
不是直接 physical field control。
103. B-05 與 SPFC 的界線
雖然先前 SPFC 探索 semantic-to-physical field compilation,
但 ANVC v0.1 不需要:
只需要普通:
- vehicle API;
- actuator;
- controller;
- cabin service。
104. Existing Technology Is Enough for v0.1
因此:
可以建立在:
- existing displays;
- existing audio;
- WebRTC;
- automotive APIs;
- central / zonal compute;
- edge AI;
- cloud AI。
105. Retrofit Path
舊車:
就可以做:
- communication continuity;
- rear work;
- entertainment;
- AI agent。
106. Retrofit Limitation
但舊車可能沒有:
- seat API;
- lighting API;
- privacy glass;
- zonal HVAC。
所以:
較低。
107. New Vehicle Path
新車可以:
- service-oriented cabin control;
- zonal architecture;
- OTA;
- multi-user;
- multi-display;
- more sensors。
108. Vehicle API Abstraction
ANVC 不應綁某 OEM。
定義:
109. CabinHAL
接口例如:
display.set_layout
audio.set_zone
seat.set_profile
climate.set_profile
lighting.set_scene
privacy.set_mode
110. HAL ≠ Raw Bus
CabinHAL 不暴露:
- arbitrary CAN;
- raw ECU write。
111. Safety Domain Boundary
更高層:
CabinHAL
只包含:
112. Safety-Critical Boundary
例如:
- steering;
- braking;
- powertrain;
- airbag;
不在普通 cabin HAL。
113. Shared Functions
有些功能:
- seat;
- window;
可能和 safety 相關。
因此需:
114. Experience Compiler Architecture
可以拆:
- Intent Parser;
- Activity Resolver;
- MRC Loader;
- Zone Evaluator;
- Capability Matcher;
- Modality Planner;
- Environment Planner;
- Safety / Privacy Validator;
- Execution Orchestrator;
- Receipt Collector。
115. Intent Parser
不直接執行。
輸出:
116. Activity Resolver
知道:
- work;
- rest;
- entertainment;
- social;
- idle。
117. MRC Loader
需要 continuity 才載入。
118. Zone Evaluator
取得:
119. Capability Matcher
看車有哪些能力。
120. Modality Planner
決定:
- visual;
- audio;
- voice;
- keyboard;
- touch。
121. Environment Planner
決定:
- light;
- seat;
- climate;
- audio scene。
122. Validator
套:
- safety;
- privacy;
- authority。
123. Orchestrator
執行 bounded commands。
124. Receipt Collector
確認:
125. Experience Plan State
定義:
126. Plan Transition
可能因:
- user;
- route;
- network;
- safety;
- fatigue;
- arrival。
127. Arrival Transition
距離目的地:
Work profile 可以:
- save;
- summarize;
- lower complexity;
- prepare exit。
128. Departure Transition
上車:
- attach Surface;
- authenticate;
- load profile;
- privacy check;
- MRC resume。
129. Ride-End Transition
shared vehicle:
- checkpoint;
- detach;
- teardown;
- local wipe;
- evidence。
130. Persistent Experience ≠ Persistent Vehicle State
使用者 profile 可持續。
但 shared car local state:
131. Reference Profile:WORK
mode: "WORK"
requirements:
visual: "high"
privacy: "high"
network: "medium_high"
preferred:
display: "private_large"
audio: "private_zone"
input:
- "voice"
- "keyboard"
environment:
seat: "upright"
light: "neutral"
fallback:
- "voice_summary"
- "offline_draft"
132. Reference Profile:REST
mode: "REST"
interruptions: "emergency_only"
visual: "off"
audio: "quiet"
seat: "reclined_if_safe"
light: "low"
network: "background_only"
133. Reference Profile:ENTERTAINMENT
mode: "ENTERTAINMENT"
visual: "immersive_if_allowed"
audio: "spatial"
work_notifications: "blocked"
134. Reference Profile:SOCIAL
mode: "SOCIAL"
display: "shared"
audio: "shared"
personal_ai: "low_intrusion"
privacy: "group_scope"
135. Reference Profile:AGENT_ONLY
mode: "AGENT_ONLY"
human_work_ui: "off"
agent_background_tasks: "allowed_by_policy"
external_side_effects: "approval_required"
arrival_review_packet: true
136. Personalization Layer
Profile 可以記:
- preferred temp;
- seat;
- light;
- audio;
- DND;
- accessibility。
但:
137. Accessibility
AI-native cabin 應支援:
- speech;
- caption;
- high contrast;
- screen reader;
- hearing assistance;
- physical accessibility;
- alternative input。
138. Accessibility Is Not Optional Luxury
如果 Mobility Hospitality 未來普及,
accessibility 應為:
139. Family Mode
兒童:
- content;
- safety;
- account;
- privacy;
不同。
AI 不應把成人 personal context 投影給 child zone。
140. Enterprise Mode
商務車:
- enterprise login;
- VPN;
- work artifact;
- private meeting;
- policy。
但:
ride end 必須 teardown。
141. Meeting Mode
可能:
- camera;
- mic;
- private audio;
- display;
- notes agent。
若 shared ride privacy 不足:
142. Call Mode
PAI Relay 可以:
但 zone:
- driver;
- passenger;
決定 interaction complexity。
143. AI as Communication Operator
車內 AI 可以:
- summarize messages;
- prepare reply;
- schedule;
- call;
- join meeting。
這承接 Series A。
144. AI as Environment Operator
另外:
- seat;
- light;
- climate;
- audio。
這是 B-05 新增。
145. AI as Safety Bypass Is Forbidden
明確:
146. AI as Experience Orchestrator
最準確:
147. Fallback Without AI
AI down:
使用者仍可:
- manual light;
- manual seat;
- normal media;
- normal navigation。
所以:
148. UX Principle:Predictable Control
AI control 必須:
- reversible;
- explainable;
- visible;
- interruptible。
149. User Override
任何非 safety cabin automation:
150. Undo
若 AI 調錯:
恢復剛才。
系統需:
151. Cabin State History
保留:
- plan;
- actuation;
- rollback;
但不必保存所有私人語音。
152. Evidence Minimization
記:
light changed
不一定記:
使用者說了完整私人句子。
153. Security Threats
T1 Prompt → Unsafe Actuation
T2 Personal Context Leakage
T3 Shared-Vehicle Residual Data
T4 Cabin/Driving Authority Confusion
T5 Malicious Device Capability
T6 Compromised Cloud Agent
T7 Fake Actuation Success
T8 Zone Misclassification
T9 Driver Distraction Escalation
T10 Persistent Microphone Overcapture
154. Threat T1 Mitigation
Intent parser 不直接 actuator。
155. T2
Data class + Zone privacy。
156. T3
Teardown contract。
157. T4
Separate agent / HAL boundary。
158. T5
Capability attestation / allowlist。
159. T6
Edge policy guard。
160. T7
Actuation receipt。
161. T8
Conservative fallback。
162. T9
B-04 ZoneDecision。
163. T10
Purpose-bound mic access。
164. ANVC MVP
第一版可以非常保守。
只做:
- rear-seat simulator;
- large display;
- voice;
- AI agent;
- audio zone simulation;
- light / seat virtual actuator;
- Continuity Runtime;
- Zone Contract。
165. MVP 不先控制真車
先:
166. Digital Twin Cabin
模擬:
- seat;
- display;
- light;
- audio;
- zone;
- motion;
- network。
167. Why Simulator First
可以驗證:
- Experience Compiler;
- MRC resume;
- ZoneDecision;
- profile;
- teardown;
不用碰 safety-critical hardware。
168. MVP Scenario 1:Desktop → Rear Cabin Work
Desktop paper
→ MRC
→ rear passenger zone
→ private display
→ voice + keyboard
→ continue
169. MVP Scenario 2:Driver Transform
same paper
→ driver zone
→ visual blocked
→ short voice cue
→ defer full resume
170. MVP Scenario 3:Rest
REST intent
→ display off
→ DND
→ low light
→ agent-only background task
171. MVP Scenario 4:Network Degradation
4K + cloud agent
→ bandwidth drops
→ local summary + offline notes
172. MVP Scenario 5:Ride End
checkpoint
→ detach
→ wipe local session
→ resume on desktop
173. Evaluation Metrics
Experience Compile Success
Zone Compliance
Actuation Success
Privacy Leakage
Continuity Loss
User Override Rate
174. Latency
speech interaction:
必須低。
NVIDIA 2026 對 on-device agentic assistant 甚至以:
作為其 production architecture 例示的 response-time challenge。
這是工程參考,不是本文 universal requirement。
175. Edge Token Throughput
同一 NVIDIA 例示:
作為 advanced local experience challenge。
這再次說明車艙 AI 已經變成 serious edge-compute workload。
176. Compute Isolation
Cabin AI workload 不應:
影響 safety-critical compute deadline。
所以:
177. Mixed-Criticality
未來 centralized platform 可能共享硬體。
但必須:
- partition;
- priority;
- isolation;
- certification boundary。
178. B-05 不定義 ADAS Safety Architecture
本文只定:
真正 ADAS / ADS safety architecture 仍屬專門領域。
179. Current Industry Status
截至 2026:
已存在:
- multiple displays;
- panoramic / windshield UI;
- zonal audio;
- spatial audio;
- multimodal LLM;
- agentic AI;
- on-device AI;
- edge-cloud pipeline;
- driver monitoring;
- centralized / zonal compute;
- conversational vehicle control。
180. 尚未成熟為統一 PMS
仍缺:
- cross-provider canonical continuity;
- portable cabin profile;
- universal Zone Contract;
- activity-aware Experience Compiler;
- shared-vehicle privacy teardown standard;
- agent / cabin / driving authority common model;
- cross-OEM experience portability。
這就是本文研究空間。
181. ANVC Maturity
ANVC-L0
Fixed infotainment。
ANVC-L1
Connected multimodal cockpit。
ANVC-L2
Context-aware AI cabin。
ANVC-L3
Zone-aware programmable cabin。
ANVC-L4
Continuity-native, portable, multi-agent programmable mobility cabin。
182. ANVC ≠ Automation Level
仍:
183. Luxury Car Can Be High ANVC, Low Automation
有司機的車:
- rear office;
- AI;
- display;
- network;
可能 ANVC 高。
184. Robotaxi Can Be Low ANVC
一台 driverless pod:
- only transport;
- tiny UI;
可能 SAE automation 高,但 ANVC 低。
185. Product Implication
因此 robotaxi differentiation 不只:
還有:
這直接進 B-06。
186. Cabin Capability Score
可定義:
187. User Experience Score
但:
高不代表:
高。
還要:
- activity match;
- comfort;
- motion;
- privacy。
188. Best Cabin Is Contextual
所以:
189. B-06 Interface
B-05 輸出:
- capability class;
- experience profiles;
- privacy class;
- zone class;
- continuity support;
- AI support。
B-06 可以拿來:
- service tier;
- pricing;
- matching;
- fleet segmentation。
190. 最重要的六條不變量
第一:
第二:
第三:
第四:
第五:
第六:
191. 結論
AI-native vehicle cabin 真正的意義不是:
車上終於有一個可以聊天的 AI。
而是:
車艙本身開始變成一個由 activity、continuity、zone、安全、隱私與使用者意圖共同編排的多模態 runtime。
現代汽車產業已經擁有大量必要 primitive:
- 多螢幕;
- windshield / panoramic interfaces;
- multi-channel / zonal audio;
- ANC;
- speech;
- multimodal LLM;
- agentic AI;
- on-device AI;
- cloud agents;
- centralized / zonal compute;
- OTA;
- context-aware personalization。
因此問題已經不是:
這些硬體能不能被放進車?
而是:
它們能不能被組成一個真正以活動與人為中心,而不是以單一 infotainment app 為中心的系統?
本文的答案是 Experience Compiler。
它讓:
成為上層。
而:
成為 backend capability。
再由:
限制。
因此:
這是一條從人的意圖一路落到物理車艙,又不讓 AI 越過安全與主權邊界的完整鏈。
B-06 接下來將回答另一個問題:
當不同車輛可以提供不同 quality 的私人/工作/休息/娛樂時空,乘客到底在買什麼?
那時交通服務就會正式從:
進入:
參考文獻與技術錨點
NVIDIA Technical Blog. How to Build In-Vehicle AI Agents with NVIDIA: From Cloud to Car. 2026-05-05.
https://developer.nvidia.com/blog/how-to-build-in-vehicle-ai-agents-with-nvidia-from-cloud-to-car/Qualcomm. Snapdragon Cockpit Platform. Current product documentation checked 2026-08-24.
https://www.qualcomm.com/automotive/solutions/cockpitQualcomm. Leapmotor and Qualcomm Debuts World's First Automotive Central Computer Powered by Snapdragon Cockpit Elite and Snapdragon Ride Elite Platforms. 2026-01-05.
https://www.qualcomm.com/news/releases/2026/01/leapmotor-and-qualcomm-debuts-world-s-first-automotive-central-cQualcomm. Qualcomm Drives the Future of Mobility with Strong Snapdragon Digital Chassis Momentum and Agentic AI for Major Global Automakers Worldwide. 2026-01-05.
https://www.qualcomm.com/news/releases/2026/01/qualcomm-drives-the-future-of-mobility-with-strong-snapdragon-diMercedes-Benz Group. Next Generation in-car intelligence. 2026-04-23.
https://group.mercedes-benz.com/technology/innovation/collaboration/liquid-ai.htmlQualcomm / Mercedes-Benz. Snapdragon Cockpit Platforms Power Smart, Intuitive AI-Driven Experiences in New All-Electric Mercedes-Benz Vehicles. 2025-09.
https://www.qualcomm.com/news/releases/2025/09/qualcomm-s-snapdragon-cockpit-platforms-power-smart--intuitive-aBMW Group. BMW Panoramic iDrive: Four Key Elements. 2025 investor / analyst materials.
https://www.bmwgroup.com/content/dam/grpw/websites/bmwgroup_com/ir/downloads/en/2025/investor-and-analyst-days-2025/BMW-Group-Investor-Analyst-Days-2025-UI-UX-Software.pdfQualcomm / BMW Group. Qualcomm Named BMW Group’s Lead Compute Silicon Provider for Digital Cockpit and Automated Driving Through the Next Decade. 2026-07-29.
https://www.qualcomm.com/news/releases/2026/07/qualcomm-named-bmw-group-s-lead-compute-silicon-provider-for-digAndroid Automotive / Android for Cars documentation cited in B-04 for multi-user / multi-display / occupant zones / UX restrictions.
Personal sound-zone literature cited in B-04 for headrest loudspeakers, acoustic contrast, ANC and in-car zonal audio.
內部依賴
B01_Programmable_Mobility_Space_General_Theory_v0.1.mdB02_Travel_Time_Reclamation_v0.1.mdB03_Mobility_Cognitive_Continuity_v0.1.mdB03_Mobility_Resume_Capsule.schema.jsonB04_Spatial_Attention_Zones_v0.1.mdB04_Spatial_Attention_Zone.schema.jsonB04_Zone_Decision.schema.json- Series A A-03 Multimodal-Native Communication
- Series A A-04 AI Communication Runtime
- Series A A-05 Adaptive Hybrid Communication Fabric
- Series A A-06 Personal Communication Sovereignty
- Series A A-07 EVEMISS Communication Continuum Architecture
文件狀態
Series: B|Programmable Mobility & Spatial Continuum
Number: B-05
Version: v0.1
Status: Research Paper / Canonical Source
Series B progress: 5/7
Overall 15-document progress: 12/15