Sensor budget matrix v1

Sensor Budget Matrix v1 (传感器预算矩阵 — 跨 embodiment SWaP-C 对照)

发布时间：2026-05-21 适用范围：6 embodiments × 8 sensor classes (manipulation/humanoid/ground/driving/aerial/marine) 核心定位：industry’s missing SWaP-C account — academic surveys never write it down, internal BoMs always do.

Status: v1.1 — opinionated draft. Backfilled to AGENTS.md 14-item dissection template 2026-05-21. Datasheet numbers marked UNVERIFIED need spec-sheet cross-check. Wedge tier: W2 (one of 5 launch docs) TL;DR: Each embodiment picks its sensor stack on a binding SWaP-C constraint academic surveys never write down. Drones bound by grams+watts, AD by range, manipulation by cost-per-cell, marine by physics. A Velodyne VLP-16 at 580g/~8W UNVERIFIED is unremarkable on a Waymo van, fatal on a 250g racing drone, over-budget by 4× on a $5k AGV. Same sensor, three verdicts.

X-Ray opening (non-expert friendly)

(a) The same VLP-16 is mandatory on Waymo’s van, fatal on a 250g racer, over-budget by 4× on a $5k AGV — one datasheet, three verdicts. (b) Each embodiment binds on a different constraint: weight+power (drones), range (AD), cost-per-cell (manipulation), physics (marine). (c) For sensor-hardware engineers, this is the BoM-level table usually rebuilt internally from scratch — academic surveys stop at “we used a RealSense D435.”

📍 研究全景时间线 (where this matrix sits)

2010 ── Velodyne HDL-64 era (KITTI). LiDAR == AD. $80k/13kg normal.
2013 ── Kinect v1 / Xtion. Indoor RGBD unlocks manipulation papers.
2015 ── MEMS IMU revolution (Bosch BMI / InvenSense). $3 IMUs unlock drones.
2017 ── 850nm RealSense D400 lineage. RGBD = "default" indoor; sun saturation
        ignored by most academic benchmarks.
2020 ── Apple iPad/iPhone Pro LiDAR (Sony dToF). Consumer SPAD normalized.
2023 ── 1550nm InGaAs SPAD wave (Hesai AT128, Innoviz, Aeva).
2025 ── Livox Mid-360 (~250g) brings LiDAR within reach of 3kg+ drones.
        ── you are here (2026) ──
?    ── Solid-state automotive LiDAR <100g? Underwater LiDAR <$50k? Open.

---

1 · SWaP-C axes per embodiment

📌 Napkin Formula (X-Ray):

verdict = sensor_cost / (embodiment_total_budget × workspace_relevance)

SWaP-C decisions are ratios, not absolute prices. A $4k LiDAR is "cheap" on an$80k AD car at 200m; “infinite” on a $300 hobby drone at 5m. Workspace relevance is the multiplier most surveys silently drop.

Embodiment	Binding constraint	Sensor share of budget
Manipulation	cost-per-cell	~5% of $25–50k arm
Humanoid	head weight + thermal	3–8%
Ground (AGV)	cost + cert	10–20% of $5–30k
Driving (L4)	range + integration miles	$50k+ on$80k vehicle
Aerial (<3kg)	weight + power	20–40% by weight
Marine (AUV)	pressure + acoustic	30–50% of $50k–1M

2 · The matrix, expanded

Cells: status · reason · number.

	RGB mono	Stereo	RGBD	LiDAR	IMU	NIR	Event	Acoustic
Manipulation	sometimes · wrist fallback	core · 5–10cm baseline	core · D435 default; 5–80cm matches workspace, 70g/$300	rare · adds nothing sub-1m	core · BMI270 ~$3	sometimes · shiny parts	rare	rare
Humanoid	core · peripheral	core · only metric primitive fitting the head	sometimes · torso, not face	sometimes · H1 ships one; 500g tolerated	core · >12 IMUs typical	rare · head thermal	research	rare
Ground (AGV)	core · cheapest primitive, $30	core · ZED-class <$500	sometimes · indoor only	core outdoor · Hokuyo $1.5k cert; Velodyne 580g/8W UNVERIFIED	core · only dead-reckoning under shelves	rare	rare	sometimes · ultrasonic <$5
Driving (AD)	core · lights/signs/lanes	sometimes · Tesla yes; Waymo no	rare · breaks past 5m; target 50–200m	core Waymo / never Tesla · 128-beam $8–80k, 1–3kg/15–30W UNVERIFIED	core · auto-grade $50–500; +FOG$5–20k tunnels	rare	research	rare · parking only
Aerial (<3kg)	core · FPV = platform’s eyes, 5–15g	sometimes · Skydio yes; racing no	rare · active illum = power suicide; §3.1	sometimes · Livox 500g/10W flies on 3kg+ only	core · only >100Hz path <1g cost; BMI270 0.5g/<5mW	rare	research · UZH RPG	sometimes · altimeter <5m
Marine (AUV)	sometimes · <5m visibility	sometimes · photogrammetry	rare · NIR absorbed <1m	rare · blue-green $200k+	core · FOG mandatory; KVH 1750 $15k/700g UNVERIFIED	rare · physics forbids	rare	core · DVL+multibeam+side-scan = the whole stack; AURELION ACOSTRA UNVERIFIED, no DOI

---

3 · The three counterintuitive cells

3.1 Why drones almost never use RGBD

A D435 is 70g/3W. On a Skydio 2+ (800g) tolerable; on a 250g cinewhoop it’s 28% of all-up weight and >10% of battery — before the projector helps, because outdoor sunlight saturates the 850nm pattern within 2m (see foundations/sensor-physics/active_nir_850nm_for_embodied_ai.md). Pay weight indoors, get nothing outdoors.

Deeper: on a 3kg quad, every 100g costs 30–60s of hover UNVERIFIED. A 500g active-illumination rig = ~2 min lost on a 12-min mission, for a sensor working only in 0.5–3m — drones fly 5–50m. The Venn diagram is empty. Pattern: mono RGB + IMU + ultrasonic <5m + (optional) downward stereo. Active depth reserved for indoor inspection drones >1.5kg.

3.2 Why AD almost never uses RGBD

Highway AD design range is 150–250m. RGBD breaks past 5m (structured light) or ~20m (ToF) because returned NIR falls as 1/r² against a 1kW/m² sunlight floor UNVERIFIED. To match LiDAR at 200m, active RGBD needs kilowatts — which is why pulsed LiDAR exists. RGBD and LiDAR aren’t competitors at AD range; they’re the same physics scaled by 100× optical power. AD depth = stereo + LiDAR + radar, RGBD only in parking-assist. Mobileye: stereo+radar+mono-ML. Waymo: LiDAR-primary.

3.3 Why manipulation almost never uses LiDAR

VLP-16: $4k/580g `UNVERIFIED`. D435:$300/70g. Manipulation workspace is a 1m³ cube. Inside it, LiDAR’s 0.1–0.4° angular resolution gives sub-cm depth — same as a $300 RGBD. Outside, you don’t care. LiDAR offers no resolution advantage in the relevant volume at 13× cost and 8× weight. Exception: mobile manipulation where base amortizes 2D LiDAR for safety — wrist stays RGBD. LiDAR’s value scales with range; manipulation’s range doesn’t.

---

4 · Cross-cutting patterns

• IMU is always core. Only question: FOG vs MEMS, decided by mission duration vs allowable drift. BMI270 (~$3, 0.5°/s bias `UNVERIFIED`) covers drones/AGVs; KVH 1750 FOG (~$15k, 0.05°/hr UNVERIFIED) covers AUVs and L4 AD.
• At least one camera is always core. Mono vs stereo is $200 + sync complexity, not physics.
• ⚡ Eureka Moment — Sensor #3 splits embodiments. IMU is universal, ≥1 camera is universal, and the third sensor (RGBD / LiDAR / sonar / nothing) is where the embodiment identity is written. Everything else follows.

5 · Hidden Assumptions

Where this matrix would break:

• Embodiment classes are fixed. Hybrid duty cycles (teleop humanoid indoor + walking outdoor) collapse here; re-classify if needed.
• Payload class dominates compute class. Sub-100g nano drones flip this — compute weight rivals sensor weight.
• Vendor SKU prices stable. Hesai/Innoviz moved 3× downward in 24 months; “$8–80k LiDAR” is a 2026 snapshot, not a law.
• Underwater visibility physics holds. Turbid coastal vs clear blue shifts “<5m visibility” by an order of magnitude.
• AD long range is a hard requirement. Geofenced low-speed AD (campus shuttle, mining truck) unlocks cheaper RGBD-class stacks.
• UNVERIFIED numbers within 2×. Order-of-magnitude is load-bearing; 5×-off vendor spec may flip individual verdicts — the pattern (Sensor #3 splits embodiments) does not.

6 · “Always” / “Never” per embodiment

Manipulation: RGBD wrist / never LiDAR. Humanoid: stereo head + multi-IMU / never active NIR on head. Ground: 2D safety LiDAR + wheel IMU / never FOG. Driving: front mono + radar + IMU / never RGBD. Aerial <3kg: IMU + mono FPV / never RGBD. Marine: FOG IMU + DVL / never RGB-only.

7 · Decision tree — embodiment + budget → starter stack

**$500 indoor AGV** → D435 ($300) + BMI270 ($3) + ultrasonic ($5) + RPi. Skip LiDAR; depth as virtual 2D scan. Ships in a weekend; won’t pass SEMI S2.

**$5k outdoor drone (1.5kg)** → Mono global-shutter ($200) + BMI270 ($3) + GPS/RTK ($300) + downward stereo ($400) + Livox Mid-360 *only if* payload >2kg (~$1k/250g UNVERIFIED). For a 250g racer: mono + IMU + baro. VLP-16 at $4k/580g is a non-starter.

**$50k AD demo car** → Auto IMU ($500) + 4–6× mono ($2k) + front stereo ($1k) + 128-beam LiDAR ($15–25k, Hesai AT128 `UNVERIFIED`) + 4× corner radar ($2k) + front long-range radar ($1.5k). LiDAR alone = half the BoM; “Tesla or Waymo school?” reduces to “spend half the budget on one device?“

8 · Falsifiable 2-year prediction

By 2028-05-21, ≥1 commercial sub-3kg drone will ship with an integrated <100g solid-state automotive-derived LiDAR (Hesai/Innoviz/Aeva lineage) at <$2k retail, flipping the aerial “LiDAR rare” cell to “sometimes” for the 1–3kg class. Falsified if no such SKU exists at that date.

9 · For the reader (per-persona)

• Manipulation engineer: stop benchmarking LiDAR; RGBD wrist + IMU is your stack — argue resolution, not range.
• Aerial engineer: every 100g = 30–60s of hover; reject active-NIR on outdoor sub-3kg platforms.
• AD engineer: RGBD is not “cheap LiDAR” — different range regime. Stereo+radar+mono-ML (Tesla) vs LiDAR-primary (Waymo).
• Marine engineer: acoustic (DVL + multibeam) is the stack; treat RGB/NIR as <5m clear-water special cases.

10 · Interview Tip

If asked “why doesn’t drone X use RGBD?” — say weight + sun saturation + range mismatch. Always SWaP-C, never “unproven technology.”

---

Cross-references

foundations/sensor-physics/active_nir_850nm_for_embodied_ai.md · embodiments/aerial/sensor-stack/ · embodiments/marine/ · deployment/hardware-selection/.

References

Sony IMX900 · Velodyne VLP-16 / Hesai AT128 / Innoviz One · Bosch BMI270 · KVH 1750 FOG · Intel RealSense D435 · AURELION ACOSTRA — all UNVERIFIED, no DOI. Skydio blog. Tesla AI Day.

Boundary

Compares sensor classes across embodiments at BoM level. Per-sensor physics: foundations/sensor-physics/. Per-embodiment integration: embodiments/*/sensor-stack/.

2026-05-21. v1.1 backfill — datasheets verified in v1.2.