> ## Documentation Index
> Fetch the complete documentation index at: https://kensou.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Sensor selection decision matrix

# 跨 Sensor 选型决策矩阵 (Cross-Sensor Selection Decision Matrix)

> **类型**: comparison (24 个 single-sensor dissection 后的元视图) — 不走 14 项 dissection 门槛
> **聚焦**: 跨 sensor 选型决策矩阵 + 决策树 + 失败模式跨 sensor 对照
> **Status**: v1 — 2026-05-22。所有具体 SWaP-C 数字 (g / W / \$) 标 `UNVERIFIED` 除非引自原 dissection

**TL;DR.** 24 个 single-sensor dissection 把波段 / 噪声 / 失败模式拆到 datasheet 一层，但**缺一个「看完之后第一步选哪个」的横向决策表**。本页补这个元视图：**§2 把 13 个 sensor class 摆在一张表上**、**§3 给 6 种典型 embodiment 的最小 viable stack**、**§4-§7 处理 single-sensor dissection 不回答的 4 个问题**（互补 / 同类怎么选 / 跨 sensor 失败模式 / 隐藏拥有成本）。最重要的判断：**没有 universal best — 只有"working range + SWaP-C + 失败模式可接受"局部最优**；250 g drone 选 16-line LiDAR、或 humanoid 选 RTK GNSS，都是 SWaP-C 决策错了一个数量级。

***

## 1 · SWaP-C 轴的回顾 — 同 sensor 在不同 embodiment 权重完全不同

| 轴           | nano drone ≤250 g | 巡检 drone 1-5 kg  | manipulator   | AD car     | humanoid         | marine AUV        |
| ----------- | ----------------- | ---------------- | ------------- | ---------- | ---------------- | ----------------- |
| **Weight**  | **\<5 g 关键**      | \<100 g          | 不敏感           | 不敏感        | \<500 g          | 浮力中性              |
| **Power**   | \<1 W (200 mAh)   | \<5 W            | wall power    | 12 V ECU   | 24 V battery     | \~50 Wh/km        |
| **Cost**    | \<\$50 BOM        | \<\$500          | \<\$2k (D435) | \<\$5k OK  | \<\$2k           | \<\$50k OK        |
| **Compute** | MCU STM32         | Jetson Orin Nano | x86 ws        | Orin AGX×多 | Orin AGX×多       | low-power MCU     |
| **Size**    | \<30 mm           | \<80 mm          | wrist-cam     | hood/roof  | head/waist 10 cm | watertight bottle |

**关键观察**：(1) nano drone 几乎只看 weight × power — 加 10 g 等于减 10% flight time；(2) AD 几乎只看 range + 失败模式，weight/cost 可吸收；(3) manipulator 几乎只看 0-1 m 米制精度；(4) marine AUV 只看声学频段，光学 / 电磁全瞎。每个 sensor 的 SWaP-C 数字回 [`README.md`](./overview.md) 24 篇 dissection。

***

## 2 · 主决策表 — 13 个 sensor class 横向

> 单位约定：range 是 typical operating range（不是 datasheet max）；Hz 是 typical update rate；Weight 是 module-level（不含 mount）；Cost 是 BOM tier（量产价位）。**所有具体数字 `UNVERIFIED` 除非链回上游 dissection。**

所有数字 `UNVERIFIED` 除非链回上游 dissection。Range = typical operating；Weight = module-level；Cost = BOM tier。

| Sensor                                                              | 原理一句话                         | Range              | Hz            | Power          | Weight     | Cost                      | 主要失败模式                           | 何时用 / 不用                               |
| ------------------------------------------------------------------- | ----------------------------- | ------------------ | ------------- | -------------- | ---------- | ------------------------- | -------------------------------- | -------------------------------------- |
| **Active-NIR 850 nm** ([↗](./active_nir_850nm_for_embodied_ai.md))  | VCSEL flood + NIR cam + BPF   | 0.1-5 m            | 30-90         | 1-5 W          | 5-30 g     | \$20-100                  | 强日光 + 镜面饱和                       | 室内/暮光 ✓ ; 正午 ✗                         |
| **ToF Phase-CW** ([↗](./tof_physics_for_embodied_ai.md))            | 调制相位测距 (Kinect/L515)          | 0.3-5 m            | 30            | 2-5 W          | 20-50 g    | \$50-300                  | wrap-around / 多径 / 阳光            | 室内 wrist depth ✓ ; >5 m ✗              |
| **ToF dToF (SPAD)**                                                 | ns 脉冲 + SPAD 计时 (iPhone)      | 0.3-10 m           | 30            | 0.5-3 W        | \<5 g      | \$5-30                    | pile-up / 弱反射                    | mobile ✓ ; 工业级 grasp ✗                 |
| **LiDAR 905 nm 机械** ([↗](./lidar_physics_905_vs_1550.md))           | Si APD + 旋转扫描                 | 30-150 m           | 10-20         | 8-20 W         | 600-1000 g | \$1-10k                   | 轴承寿命 / 雨雪 Mie                    | 测绘 ✓ ; 量产 AD ✗                         |
| **LiDAR 905 nm 半固态**                                                | MEMS/OPA + Si APD (AT128)     | 100-200 m          | 10            | 8-15 W         | 300-700 g  | \$1-4k                    | FOV 窄 / 反射率敏感                    | 乘用车 AD ✓ ; nano drone ✗                |
| **LiDAR 1550 nm FMCW**                                              | InGaAs SPAD + chirp 多普勒       | 200-500 m          | 10            | 15-30 W        | 800-2000 g | \$5-20k                   | 单像素 50-100× Si cost              | 卡车/远距 ✓ ; cost-bound ✗                 |
| **RGB Camera GS** ([↗](./rgb_camera_imaging_pipeline.md))           | CMOS + Bayer + global shutter | passive            | 30-120        | 0.5-2 W        | 5-30 g     | \$5-100                   | dazzle / textureless / blur      | 几乎所有 embodiment ✓ ; 无光/烟雾/水下 ✗         |
| **RGB RS** ([↗](./rolling_vs_global_shutter.md))                    | rolling shutter readout       | passive            | 30-60         | 同上             | 同上         | 更便宜                       | jello / fast yaw 几何崩             | static ✓ ; drone racing ✗              |
| **Stereo (passive)** ([↗](./stereo_camera_geometry_physics.md))     | `Z=fB/d` 三角                   | baseline×(10-100)  | 30-60         | 2-5 W          | 50-150 g   | \$200-2k                  | textureless / 标定漂 / 曝光不对称        | manipulation/drone 避障 ✓ ; >100 m/白墙 ✗  |
| **Event Camera (DVS)** ([↗](./event_camera_dvs_physics.md))         | 异步 per-pixel log-intensity    | passive, \~1 µs    | 异步            | 0.1-1 W        | 30-100 g   | \$1-3k (Prophesee)        | 静态没事件 / 生态嫩                      | HDR/高速/低光 ✓ ; 量产认证 ✗                   |
| **IMU MEMS** ([↗](./imu_physics_and_noise_model.md))                | Coriolis MEMS + 加速度计          | proprio            | 100-1000      | \<0.1 W        | \<5 g      | \$1-10 (BMI270)           | bias drift / 振动 / 温漂；60 s ≈ 18 m | 全 embodiment baseline ✓ ; 长时单用 ✗       |
| **IMU tactical**                                                    | Honeywell HG4930 级            | proprio            | 200-2000      | 1-3 W          | 100-200 g  | \$3-10k                   | 振动 / 寿命                          | 巡航 / AGV ✓ ; 消费品 ✗                     |
| **IMU FOG**                                                         | Sagnac 光纤干涉, BI 0.05°/hr      | proprio            | 100-200       | 3-10 W         | 500-1500 g | \$10k-100k                | 体积大 / 需温稳                        | 真值 / 潜艇 ✓ ; weight-bound ✗             |
| **GNSS L1 single** ([↗](./gnss_multi_constellation_rtk.md))         | 单频 code phase                 | global, 5-10 m     | 1-10          | 0.1-0.3 W      | \<10 g     | \$10-30                   | 峡谷/多径/电离层                        | 户外 baseline ✓ ; 室内 ✗                   |
| **GNSS multi-band RTK**                                             | 双频 + base + carrier (F9P)     | 1-5 cm             | 5-20          | 0.5-2 W        | 30-100 g   | \$200-1k                  | base 失联即回退                       | 测绘/commercial drone ✓ ; 室内 ✗           |
| **GNSS PPP**                                                        | 多频 + precise products         | \~10 cm, 收敛 30 min | 1             | 0.5-2 W        | \~RTK      | \$1k-5k                   | 收敛慢                              | marine/静态 ✓ ; 实时机动 ✗                   |
| **Magnetometer** ([↗](./magnetometer_geomagnetic_field.md))         | Hall/fluxgate 测地磁             | yaw global         | 10-100        | \<0.05 W       | \<1 g      | \$1-30                    | 硬/软铁；50 A ESC vs 50 µT           | drone yaw ✓ ; 钢梁/电流附近 ✗                |
| **mmWave 77 GHz AD** ([↗](./mmwave_radar_physics_for_ad.md))        | FMCW + MIMO 阵列                | 5-300 m            | 10-30         | 5-15 W         | 200-500 g  | \$500-3k 模组               | 角分辨率 \~0.5° / 金属混叠               | AD/浓雾穿透 ✓ ; 室内 dense ✗                 |
| **mmWave 24 GHz** ([↗](./24ghz_doppler_radar_motion.md))            | K-band CW/FMCW                | 0.5-20 m           | 10-50         | 0.3-2 W        | 5-30 g     | \$5-50                    | 角分辨率粗 / 频段限带宽                    | smart home / 廉价防撞 ✓ ; AD ✗             |
| **UWB** ([↗](./uwb_ultra_wideband_positioning.md))                  | 500 MHz 脉冲 + DS-TWR           | 0.1-50 m, σ ≈ 3 cm | 10-100        | RX 150 mW peak | \<5 g      | $3-8 chip / $30-50 anchor | 多径 / NLOS bias / 需 anchor        | 室内 cm/swarm ✓ ; 户外大场景 ✗                |
| **Thermal IR LWIR** ([↗](./thermal_ir_lwir_8_14um.md))              | microbolometer 被动黑体           | 0-100 m            | 9-60 (export) | 0.5-2 W        | 5-100 g    | \$100-3k (Boson+)         | 不透玻璃 / 热饱和 / FFC 1 s 盲           | 夜视/烟雾/救援 ✓ ; 透玻璃 ✗                     |
| **Barometer** ([↗](./barometer_pressure_altimetry.md))              | MEMS 压力 + 温补                  | -500 to 9000 m     | 10-100        | \<0.01 W       | \<0.5 g    | \$2-5 (BMP388)            | 温漂 / 阵风 / HVAC                   | drone vertical mid-term ✓ ; 高动态\<1 s ✗ |
| **Ultrasonic** ([↗](./ultrasonic_acoustic_physics_for_robotics.md)) | 40 kHz airborne ToF           | 0.05-4 m           | 10-40         | 0.3-1 W        | 5-20 g     | \$1-30                    | 多径 / 软材吸收 / 风漂                   | drone 起降/USS ✓ ; >5 m/风大 ✗             |

23 行覆盖 12 个**主** sensor class（active-NIR / ToF / LiDAR 3 架构 / RGB GS+RS / stereo / event / IMU 3 等级 / GNSS 3 等级 / magnetometer / mmWave 2 频段 / UWB / thermal）+ drone 专用 barometer + ultrasonic。完整 SWaP-C 数字回各 dissection 与 [README.md](./overview.md)。

***

## 3 · 决策树 (by use case) — 6 个典型 embodiment

> 每个 use case 给"最小 viable stack"（去掉 1 项就有 spec 跑不到）+ 可选增强 + **明确点名不要选的 sensor**。

### 3.1 · 250 g 微型 drone 室内 SLAM (Crazyflie / Skydio nano)

**约束**: weight \<5 g/sensor, power \<1 W, cost \<\$50。**Stack**: IMU MEMS (BMI270) + monochrome GS camera + 单点 NIR ToF / optical flow PMW3901 + barometer (BMP388)。**不要**: ✗ LiDAR (≥300 g 超 weight budget 60×) / ✗ RTK GNSS (室内无信号 + 30 g) / ✗ thermal IR (cost + FFC 不划算)。
参考: [imu](./imu_physics_and_noise_model.md) + [optical\_flow](./optical_flow_sensor_pmw3901.md) + [rolling\_vs\_global\_shutter](./rolling_vs_global_shutter.md) + [`embodiments/aerial/sensor-stack/`](../../embodiments/aerial/sensor-stack/)

### 3.2 · 1.5 kg 巡检 drone 户外 (DJI M300 / Skydio X10)

**约束**: weight \<100 g/sensor, power \<5 W, cost \<\$500。**Stack**: IMU (ICM-42688) + GS stereo (D435) + RTK GNSS (F9P) + magnetometer (RM3100, 远 ESC) + barometer。**可选**: 16-line LiDAR (\~600 g) for terrain following / 77 GHz mmWave (\~50 g) for BVLOS 浓雾备份。
参考: [stereo](./stereo_camera_geometry_physics.md) + [gnss](./gnss_multi_constellation_rtk.md) + [magnetometer](./magnetometer_geomagnetic_field.md) + [`embodiments/aerial/long-range-slam/`](../../embodiments/aerial/long-range-slam/) + [`embodiments/aerial/obstacle-avoidance/`](../../embodiments/aerial/obstacle-avoidance/)

### 3.3 · 室内 manipulation wrist (Franka / UR / xArm)

**约束**: 0-1 m 米制精度 \<few mm。**Stack**: active stereo RGBD (D435) + RGB GS + force-torque (非本目录)。**可选**: ToF (L515, 已 EOL) for dense / event camera for 高速 dynamic grasp / Azure Kinect for 离线 calibration。**不要**: ✗ LiDAR (overkill + sparse) / ✗ GNSS / mmwave / UWB (室内有 frame)。
参考: [tof](./tof_physics_for_embodied_ai.md) + [stereo](./stereo_camera_geometry_physics.md) + [active\_nir](./active_nir_850nm_for_embodied_ai.md) + [`embodiments/manipulation/3d_feature_cloud_representations.md`](../../embodiments/manipulation/3d_feature_cloud_representations.md)

### 3.4 · AD-class 高速车 (Waymo / Mobileye)

**约束**: range 200+ m + 失败模式覆盖 + 冗余。**Stack**: 8-12 颗 GS camera + 4D mmWave (Arbe Phoenix) + 1-3 颗 905 nm 半固态 LiDAR + GNSS multi-band + tactical IMU loose coupling。**可选**: 1550 nm FMCW (卡车 / 远距) / thermal IR (夜视行人，但 Tesla 至今不加)。**Tesla doctrinal 例外**: ✗ LiDAR — vision + radar 路线，详见 [`embodiments/driving/waymo_vs_tesla_doctrinal_split.md`](../../embodiments/driving/waymo_vs_tesla_doctrinal_split.md)。
参考: [lidar](./lidar_physics_905_vs_1550.md) + [mmwave](./mmwave_radar_physics_for_ad.md) + [thermal\_ir](./thermal_ir_lwir_8_14um.md) + [`embodiments/driving/`](../../embodiments/driving/)

### 3.5 · Marine surface / underwater AUV

**物理约束**: 水下光 + 电磁 (>500 MHz) 全衰减。**Surface USV**: GNSS RTK + tactical IMU + RGB。**Underwater AUV**: DVL + tactical/FOG IMU + multibeam/side-scan sonar + USBL/LBL acoustic positioning。**不要 (物理)**: ✗ LiDAR (水中 \<10 m 衰减完) / ✗ mmWave / UWB (水导体) / ✗ thermal IR (表面下完全无用)。
参考: [underwater\_sonar](./underwater_sonar_physics.md) + [`embodiments/marine/sensor_stack_underwater.md`](../../embodiments/marine/sensor_stack_underwater.md) + [`embodiments/marine/underwater_slam_dvl_sonar.md`](../../embodiments/marine/underwater_slam_dvl_sonar.md)

### 3.6 · Humanoid (Unitree H1 / Figure 02 / 1X NEO)

**约束**: 多 viewpoint (head + waist + wrist), 24 V battery。**Stack**: head stereo RGBD (D455) + waist downward stereo (脚下地形) + wrist RGBD/ToF (D405/L515) + torso IMU + foot pressure。**可选**: Livox Mid-360 16-line LiDAR (\~265 g) for 大场景 SLAM (Unitree 路线) / 24 GHz mmWave for presence。
参考: [`embodiments/humanoid-legged/whole_body_spatial_perception.md`](../../embodiments/humanoid-legged/whole_body_spatial_perception.md) + [`embodiments/humanoid-legged/unitree_h1_vs_figure_vs_1x.md`](../../embodiments/humanoid-legged/unitree_h1_vs_figure_vs_1x.md)

***

## 4 · 互补组合 — 为什么"加一颗"是物理必然

任何 single-sensor 在某个轴上都瞎；下面 4 个组合**不可裁剪**：

* **IMU + Camera = minimum viable VIO**。IMU 单独 60 s 漂 \~18 m + 3° 转角 (`UNVERIFIED`，[imu](./imu_physics_and_noise_model.md) §3)；camera 单独 scale ambiguous / baseline 限远端。**IMU 给 metric scale + 高频 propagation；camera 给 absolute drift correction** — EKF/MSCKF/VINS 都是这个数学。"VIO 是 spatial intelligence minimum unit" 的物理根据，不是"算法选择"。
* **GNSS + IMU + Visual/LiDAR = 户外完整 stack**。GNSS 单独城市峡谷 / 桥下 / 室内全瞎；GNSS+IMU loose-coupling dropout 1 s OK 但 60 s 漂 18 m；加 visual/LiDAR 后 dropout 期间 VIO 接管，camera/LiDAR 提供 absolute features。Commercial drone / robotaxi / agriculture 事实标准。
* **mmWave + Camera = 天气韧性**。Camera 雨/雾/扬尘衰减 50-80%（[lidar](./lidar_physics_905_vs_1550.md) §6）；77 GHz mmWave 穿透 30-100× + native velocity 但角分辨率粗（[mmwave](./mmwave_radar_physics_for_ad.md) ⚡ Eureka）。Tesla 重新加 radar 的根因不是 cost；Mobileye / Waymo / Bosch ARS540 都用这套。
* **Magnetometer + Optical Flow + Barometer = nano drone 室内三件套**。没 GNSS 时，mag 给 yaw（远 ESC）、flow 给 velocity（地面有 texture）、baro 给 altitude mid-term（避 HVAC 阵风）— 三者独立全弱，合一加 IMU 撑 Crazyflie 室内 hover。

***

## 5 · 同类内部选型 — 关键分歧

### 5.1 LiDAR: 905 nm vs 1550 nm vs 架构

| 维度                  | 905 nm Si APD                  | 1550 nm InGaAs SPAD         |
| ------------------- | ------------------------------ | --------------------------- |
| QE @ λ              | \~30% `UNVERIFIED`             | \~20-30% `UNVERIFIED`       |
| Eye-safe peak power | \~5 mW                         | \~5 W (\~1000×)             |
| 像素 cost             | baseline                       | 50-100×                     |
| 雨雪环境                | Mie scatter 主导（差异小）            | peak power headroom 给"看穿"裕量 |
| 量产路线                | Hesai AT128 / Innoviz / Ouster | Luminar Iris / Aeva         |
| 决策准则                | **量产乘用车 / 成本敏感**               | **远距 / 卡车 / 高速速度场**         |

机械 vs 固态 vs FMCW 的分歧：**机械**寿命有限（轴承）但 360° FOV，**MEMS 半固态**寿命好但视场角窄需 stitch，**FMCW**直接拿 velocity 但单像素贵 — 见 [lidar](./lidar_physics_905_vs_1550.md) §7。

### 5.2 IMU: MEMS / tactical / FOG — Allan plot 的"等级"

| Grade                  | Bias instability      | Weight     | Cost       | 60 s drift (位置) `UNVERIFIED` |
| ---------------------- | --------------------- | ---------- | ---------- | ---------------------------- |
| Consumer MEMS (BMI270) | 0.5°/s `UNVERIFIED`   | \<5 g      | \$1-10     | \~18 m                       |
| Tactical (HG4930)      | 0.05°/hr `UNVERIFIED` | 100-200 g  | \$3-10k    | \~几 m                        |
| FOG (KVH 1750)         | 0.05°/hr `UNVERIFIED` | 500-1500 g | \$10k-100k | \<1 m                        |

数字回 [imu](./imu_physics_and_noise_model.md) §3。**关键 trade-off**：consumer MEMS 已经能跑 manipulation / 消费 drone；tactical 在 GNSS-denied 巡航中是 must-have；FOG 在自动驾驶量产车上完全不现实（weight × cost 都炸）— 只用在 robotaxi 数据采集车做 ground truth。

### 5.3 GNSS: single / RTK / PPP

| 等级             | 精度      | 收敛时间                  | base station?           | 适用                         |
| -------------- | ------- | --------------------- | ----------------------- | -------------------------- |
| L1 single      | 5-10 m  | \<1 s                 | no                      | 消费 drone / 农业 baseline     |
| Multi-band RTK | 1-5 cm  | 5-30 s + carrier lock | **yes**                 | 测绘 / 建筑 / commercial drone |
| PPP            | \~10 cm | 30 min                | no (用 precise products) | marine / 静态 / 农业           |

详见 [gnss](./gnss_multi_constellation_rtk.md) §5。**RTK 需要 base 在 10-30 km 内**，跨海 / 跨州时退化到 PPP。

***

## 6 · 失败模式跨 sensor 对照 — 同一扰动崩多少个

> ⚠️ = 中等退化, ❌ = 完全失效, ✅ = 鲁棒

| 扰动                            |        RGB       |       Stereo      |       Active-NIR       | ToF |  905 nm LiDAR |         mmWave        |   UWB  |        IMU        |     GNSS     |    Mag    |  Thermal IR  | Sonar |
| ----------------------------- | :--------------: | :---------------: | :--------------------: | :-: | :-----------: | :-------------------: | :----: | :---------------: | :----------: | :-------: | :----------: | :---: |
| **正午阳光 dazzle**               |         ❌        |         ❌         |       ⚠️ (BPF 救)       |  ⚠️ |  ⚠️ (ambient) |           ✅           |    ✅   |         ✅         |       ✅      |     ✅     |      ⚠️      |   ✅   |
| **雨 10 mm/hr**                |        ⚠️        |         ⚠️        |           ⚠️           |  ⚠️ | ❌ Mie scatter |      ✅ 衰减 \<3 dB      |   ⚠️   |         ✅         | ⚠️ multipath |     ✅     |      ⚠️      |  n/a  |
| **浓雾 (能见 \<50 m)**            |         ❌        |         ❌         |            ❌           |  ❌  |       ❌       | ✅ (77 GHz 穿透 30-100×) |   ⚠️   |         ✅         |      ⚠️      |     ✅     |    ⚠️ (湿)    |  n/a  |
| **textureless 白墙 / 雪地**       |        ⚠️        |   ❌ disparity 没了  |   ✅ (project pattern)  |  ✅  |       ✅       |           ✅           |    ✅   |         ✅         |       ✅      |     ✅     |   ⚠️ (温度均匀)  |  n/a  |
| **镜面反射 / 玻璃**                 |        ⚠️        |         ❌         |    ❌ (specular 不返回)    |  ❌  |       ❌       |           ✅           |    ✅   |         ✅         |       ✅      |     ✅     |   ❌ (不透玻璃)   |  n/a  |
| **GNSS spoofing / 城市峡谷**      |         ✅        |         ✅         |            ✅           |  ✅  |       ✅       |           ✅           |    ✅   |         ✅         |       ❌      |     ✅     |       ✅      |  n/a  |
| **大铁结构 / 高电流 (50 A ESC)**     |         ✅        |         ✅         |            ✅           |  ✅  |       ✅       |           ✅           |    ✅   |         ✅         |      ⚠️      | ❌ 硬铁 + 软铁 |       ✅      |  n/a  |
| **UWB 多径 / NLOS**             |         ✅        |         ✅         |            ✅           |  ✅  |       ✅       |           ⚠️          | ❌ bias |         ✅         |       ✅      |     ✅     |       ✅      |  n/a  |
| **桨叶振动 (200-2000 Hz)**        | ⚠️ (motion blur) |         ⚠️        |           ⚠️           |  ⚠️ |       ⚠️      |           ✅           |    ✅   | ❌ aliasing + bias |       ✅      |     ⚠️    |      ⚠️      |  n/a  |
| **>30°C 热漂 / 太阳曝晒**           |         ✅        | ⚠️ baseline drift |  ⚠️ VCSEL 漂 0.06 nm/°C |  ⚠️ |       ⚠️      |           ✅           |    ✅   |    ⚠️ bias temp   |       ✅      |     ⚠️    | ⚠️ FFC pause |   ✅   |
| **水下 / 透过水面**                 |     ❌ \<10 m     |         ❌         |            ❌           |  ❌  |       ❌       |           ❌           |    ❌   |         ✅         |       ❌      |     ⚠️    |       ❌      |  ✅ 唯一 |
| **强 fluorescent 闪烁 (400 Hz)** |   ⚠️ (banding)   |         ⚠️        | ⚠️ (Phase-CW 拍频 ghost) |  ❌  |       ✅       |           ✅           |    ✅   |         ✅         |       ✅      |     ✅     |       ✅      |   ✅   |

**5 个跨 sensor 的"协同崩"模式**（设计时务必加冗余）：

1. **光学 + 镜面反射**：RGB / stereo / NIR / ToF / LiDAR 全瘫 — 唯一答案是 polarization 或 mmWave；
2. **光学 + 浓雾**：所有光学全瘫 — 唯一答案是 77 GHz mmWave；
3. **声学 + 强混响 / 软材料**：sonar / ultrasonic 同时崩 — 水下 UAV 设计要避开石壁直角；
4. **电磁 + 室内 + 干扰**：GNSS / WiFi RTT / mag 都不靠谱 — 室内必须 UWB or VIO；
5. **慣性 + 桨叶振动**：MEMS IMU 在 200-2000 Hz aliasing — drone 必须橡胶减振 + 200+ Hz IMU。

***

## 7 · 真实拥有成本 — BOM 单价是冰山一角

24 个 dissection 给的是 BOM 单价；生产环境真实成本至少 **3-10×**。隐藏成本: (1) **Calibration toolchain** (Kalibr / IMU-cam / multi-cam) — 1 工程师月 + 标定室；(2) **ROS driver maintenance** — 每代固件 1-2 周回归；(3) **Factory line test fixture** — $5k-50k 检具 + 工时；(4) **Per-unit calibration** 5-30 min/unit；(5) **Field re-cal** (drone stereo ~0.024 m/m drift `UNVERIFIED`)；(6) **认证 / 监管** (FCC / IEC 60825 / FAA / CE) — $30k-300k + 6-12 月；(7) **出口管制** (中国 LiDAR / FOG ITAR)。

**经验法则**: D435 BOM $300 → 量产线 ~$1000；车规 LiDAR $1k → 含认证维保 ~$5k；FOG IMU $20k → 含 ITAR licensing ~$50k+。详细见 [`deployment/hardware-selection/`](../../deployment/hardware-selection/)。

***

## 8 · 这张矩阵不能告诉你的事

(1) **Form factor / mechanical mounting** — connector 5 g 可能就让 sensor 嵌不进 250 g drone；(2) **Vendor lock-in** — RealSense / Livox SDK 迁移成本不在 BOM；(3) **国贸 / 出口管制** — 中国 LiDAR 受限 / FOG ITAR，物理能买 ≠ 合规能用；(4) **Regulatory** — 事件相机 / 1550 nm LiDAR 乘用车 functional safety 认证仍在演化；(5) **供应链 lead time** — Sony 高端 IMX / FLIR Boson 在 2024-25 出现 12-24 月 lead time；(6) **二级市场** — 工业 LiDAR 二手不存在，故障即报废；(7) **算法生态** — DVS 算法库 vs 传统 RGB pipeline 差 5-10 年成熟度；(8) **新 paradigm** — 2025-11 DA 3 / VGGT-Ω / MapAnything 把单目/stereo/multi-view 边界打散（见 [`../depth-foundation/depth_models_comparison.md`](../depth-foundation/depth_models_comparison.md)），可能让 camera-only 在某些场景重新可用。

**正确读法**：作为**第一遍 filter** 淘汰明显不合的；剩 2-3 候选回 single-sensor dissection 查 worked example；最后 vendor EVK 实测再下单。**没有 matrix 替代真机测试**。

***

## Cross-references — 24 个 single-sensor dissection 的常用组合

**Aerial 主推 5 件套**：

* [imu\_physics\_and\_noise\_model](./imu_physics_and_noise_model.md) — 必备 baseline
* [gnss\_multi\_constellation\_rtk](./gnss_multi_constellation_rtk.md) — 户外 absolute
* [stereo\_camera\_geometry\_physics](./stereo_camera_geometry_physics.md) — 视觉避障
* [magnetometer\_geomagnetic\_field](./magnetometer_geomagnetic_field.md) — yaw 唯一绝对参考
* [barometer\_pressure\_altimetry](./barometer_pressure_altimetry.md) — vertical mid-term
* → 集成实战见 [`embodiments/aerial/sensor-stack/`](../../embodiments/aerial/sensor-stack/) + [`embodiments/aerial/vio/`](../../embodiments/aerial/vio/) + [`embodiments/aerial/long-range-slam/`](../../embodiments/aerial/long-range-slam/)

**AD 主推 7 件套**：

* [rgb\_camera\_imaging\_pipeline](./rgb_camera_imaging_pipeline.md) — BEV 主输入
* [rolling\_vs\_global\_shutter](./rolling_vs_global_shutter.md) — 高速场景必 GS
* [lidar\_physics\_905\_vs\_1550](./lidar_physics_905_vs_1550.md) — geometric truth
* [mmwave\_radar\_physics\_for\_ad](./mmwave_radar_physics_for_ad.md) — 天气韧性
* [gnss\_multi\_constellation\_rtk](./gnss_multi_constellation_rtk.md) — loose-coupling localization
* [imu\_physics\_and\_noise\_model](./imu_physics_and_noise_model.md) — tactical-grade
* [thermal\_ir\_lwir\_8\_14um](./thermal_ir_lwir_8_14um.md) — 夜视 / 行人（可选，Tesla 例外）
* → 集成实战见 [`embodiments/driving/`](../../embodiments/driving/) + [`embodiments/driving/waymo_vs_tesla_doctrinal_split.md`](../../embodiments/driving/waymo_vs_tesla_doctrinal_split.md)

**Manipulation 主推 4 件套**：

* [stereo\_camera\_geometry\_physics](./stereo_camera_geometry_physics.md) — D435 wrist
* [tof\_physics\_for\_embodied\_ai](./tof_physics_for_embodied_ai.md) — L515 dense depth
* [active\_nir\_850nm\_for\_embodied\_ai](./active_nir_850nm_for_embodied_ai.md) — 暗光 / textureless
* [rgb\_camera\_imaging\_pipeline](./rgb_camera_imaging_pipeline.md) — color + features
* → 集成实战见 [`embodiments/manipulation/`](../../embodiments/manipulation/)

**Marine 主推 3 件套**：

* [underwater\_sonar\_physics](./underwater_sonar_physics.md) — 水下唯一可信
* [gnss\_multi\_constellation\_rtk](./gnss_multi_constellation_rtk.md) — 水面 RTK
* [imu\_physics\_and\_noise\_model](./imu_physics_and_noise_model.md) — tactical / FOG
* → 集成实战见 [`embodiments/marine/sensor_stack_underwater.md`](../../embodiments/marine/sensor_stack_underwater.md) + [`embodiments/marine/underwater_slam_dvl_sonar.md`](../../embodiments/marine/underwater_slam_dvl_sonar.md)

**通用噪声框架**（跨所有 embodiment）：

* [sensor\_noise\_modeling\_allan\_variance](./sensor_noise_modeling_allan_variance.md) — IMU / camera / LiDAR / radar 统一到一个 Allan plot
* [event\_camera\_dvs\_physics](./event_camera_dvs_physics.md) — 跨 embodiment 的新 sensor 范式
* [polarization\_sensing\_for\_3d](./polarization_sensing_for_3d.md) — 玻璃 / 透明物体物理解

**跨 embodiment BOM 矩阵** (本文 + 6 embodiment 的应用层综合) → [`../../crossing/sensor-stack-matrix/`](../../crossing/sensor-stack-matrix/)

***

## Boundary

* **本文档位置**：comparison（24 个 single-sensor dissection 后的元视图），不走 14 项 dissection 门槛 — 核心是横向对比表 + 决策树
* **per-sensor 物理深拆**（QE / Allan plot / 失败模式细节）→ 回各 single-sensor dissection
* **per-embodiment 集成实战**（calibration / 安装位置 / 同步）→ [`embodiments/<emb>/sensor-stack/`](../../embodiments/)
* **跨 embodiment BOM SWaP-C 矩阵**（6 embodiment × 8 sensor class）→ [`crossing/sensor-stack-matrix/`](../../crossing/sensor-stack-matrix/)
* **vendor / SKU 选型 / 供应链**（D435 vs V100 vs Boson）→ [`deployment/hardware-selection/`](../../deployment/hardware-selection/)
* **Sensor + ML 融合算法**（RAFT-Stereo / Polarization-NeRF）→ [`foundations/feed-forward-3d/`](../feed-forward-3d/) + [`foundations/depth-foundation/`](../depth-foundation/)

***

[← Back to Sensor Physics](./overview.md)
