Doc 5 · Teach It to Aim — Spatial Intelligence for Coordinated Spotlights¶

Engineered Lighting prototype series · July 2026 Your V-JEPA model answers what is happening (reading, cooking, movie). This document answers where it's happening and what pan/tilt angles put light there — for multiple fixtures, coordinating. Constraints held: all video local; central RTX 6000 box runs perception; one camera per room; assisted room scan OK; HA/MQTT ecosystem; fixture form factor deprioritized (floor lamps fine).

Hardware assumption, stated plainly: the phases here run on the already-owned RTX 6000 Blackwell workstation. That card matters for the full vision (many rooms + V-JEPA); Phase 0's single-camera tracking runs fine on any recent CUDA GPU with ~12 GB+, so the first demo doesn't wait on big hardware.

Bill of Materials — per room + one-time, ~$120–320/room¶

#	Part	Qty	Est.	Where	Notes
1	PoE camera — standard rooms: Amcrest IP5M-T1179EW-AI-V3 (5 MP, 133° lens)	1/room	$80	amcrest.com	Phase 0. The room's eyes: a wide fixed lens covering the whole room from a corner, streaming standard RTSP video the GPU box ingests directly. PoE = one cable carries power and data, and wired video is what keeps a many-camera home stable and lets the camera network run fully internet-isolated
1b	— priority rooms: Loryta/Dahua IPC-T549M-ALED-S3 (full-color night vision)	opt.	$185	empiretech01.com	Phase 1+, optional. Night matters: IR cameras output grayscale, which degrades RGB-trained models
2	PoE switch (8–16 port) + Cat6 runs	1/home	$150–400	Amazon	Phase 1+ (Phase 0's single camera can run off any PoE injector). Put cameras on their own VLAN with no internet route — privacy by construction
3	ChArUco calibration board (print + mount on rigid backing)	1	~$10	print it (calib.io generator)	Phase 1. For registering each camera to the room scan. Rigid and flat matters
4	iPhone/iPad Pro with LiDAR (borrow one) + Polycam or Scaniverse app	1	free–$	App Store	Phase 1. One-time room scan, ~1–3 cm accuracy. Use Space Mode (LiDAR), not the auto-floorplan mode
5	Optional, later: one stereo depth camera — Orbbec Gemini 335L ($359) or Luxonis OAK-D Pro PoE ($579)	0–1	$359 / $579	store.orbbec.com, shop.luxonis.com	Tier 2, later. True per-pixel 3D for the specific rooms where ground-plane geometry runs out (reclined people, heavy clutter) — buy after Tier-0 tracking measurably misses, not before (see "the depth question")
—	Already owned	—	—	—	RTX 6000 Blackwell box (runs all perception), the Doc 3/4 bench rig (the actuator being aimed), Home Assistant + MQTT (the message plumbing)

Concepts (plain English)¶

Frames & transforms: every position lives in some coordinate system ("frame") — the room, a camera, a fixture. A transform converts between them. The whole aiming stack is: target in room frame → transform to fixture frame → two angles.
Intrinsics / extrinsics: a camera's intrinsics describe its lens (how pixels map to directions); extrinsics are where it sits in the room. Calibrate both once, then every pixel becomes a 3D ray.
Ground-plane trick (homography): a person's feet touch the floor, and we know where the floor is. So one ordinary camera gives their 3D position: pixel → ray → intersect floor. This is the finding that deletes depth sensors from the plan.
Keypoints / pose: models that find body joints (ankles, wrists, head) in an image. Body-scale keypoints work across a room; finger-scale does not (hands are 20–40 pixels at 4–8 m — too small, verified 2026 research).
Tracking ID: detection says "a person is here this frame"; tracking says "it's the same person as last frame." Needed so beams follow individuals, not flicker between people.
Open-vocabulary detection: models you query with words ("book") instead of fixed class lists. Used to confirm a book is present, not to measure it precisely.
Latency vs. update rate: latency = how stale each measurement is; update rate = how often they arrive. Smooth following needs ~15–30 updates/second and prediction to cover the staleness (below).
One Euro filter: the standard smoothing trick for noisy tracking — heavy smoothing when still (kills jitter), light smoothing when moving (kills lag). Cheap, three lines of code.
Fast path / slow path: two loops sharing the cameras. Fast: person tracking at 15–30 Hz → aim. Slow: video clips → V-JEPA → "they're reading" (arrives seconds later, changes policy, not coordinates). The fast path never waits for the slow one.
No-go cone: a computed exclusion zone around each tracked head that beams must never enter — borrowed from automotive adaptive headlights, which mask glare zones around oncoming drivers in exactly this way.
Named tools, one line each: YOLO26 — the current fast, standard "find the people in this frame" model. ByteTrack — the standard "it's the same person as last frame" ID-keeper. Frigate — an open-source camera recorder/reviewer that runs on the GPU box (used for recording and clips, never in the aiming loop). DeepStream — NVIDIA's toolkit for running vision models over many camera streams at once (an optional productized shortcut). go2rtc — a small program that re-serves camera streams with minimal delay.

1. What we're building (the six findings)¶

No stereo needed to start. Ground-plane localization delivers 4–19 cm accuracy in room-scale tests (4–13 cm standing, 16–19 cm walking) — comfortably inside the real beam, which the optics research pinned down: the tightest off-the-shelf 3-up optic (Carclo 10507, ~16°) throws a 0.42 m spot at 1.5 m, growing to ~0.85 m at 3 m. NVIDIA productized this pattern (DeepStream SV3DT/MV3DT, auto-calibration, 645–1,225 FPS aggregate on this GPU class). Seated people (feet unusable): the room model supplies the couch height, torso keypoints supply the rest. Stereo is a targeted later upgrade.
Beam self-calibration is the differentiator. No commercial system — including $30k zactrack/BlackTrax — auto-calibrates fixtures; all use a manual jog onto known points. But the recipe exists (a 2025 patent describes it; telescope pointing models supply the math with ~170× margin over our 1° need): sweep the beam, let the camera watch where it lands, fit the model. A fixture that calibrates itself in ~2 minutes beats the industry's shipped state of the art. (Patent flag: US 12,190,542.)
The follow-me loop is real but unforgiving. 100–300 ms end-to-end is achievable only on a bespoke path (camera → own tracker → fixture → CAN; the fixture link is MQTT on the bench, graduating to a direct ESPHome native-API action called via aioesphomeapi in production — one hop, typed floats, spec in Doc 6 §1). Convenience shortcuts cost seconds (Frigate events: 6–10 s; HA automations: documented 4–20 s traps). And prediction is structural, not polish: at 300 ms latency a walker moves 42 cm — comparable to the beam's half-width — so the aim point must lead the track by the measured latency. Geometry helps too, and for free: the fixed 10507 optic naturally widens with throw (~0.85 m at 3 m), so the beam is at its most forgiving exactly where tracking error and lag are largest; the tight-spot regime only exists at short throw, where tracking is at its best. (No variable-beam hardware needed — if beam-width experimentation matters later, the Carclo optics snap-swap in seconds, and a motorized-zoom path exists: Khatod's 45 mm slide-zoom, 11–45°, at the cost of redesigning to a single-die emitter.)
Fixture speed requirement (computed): a person passing 1–1.5 m from a fixture's plumb line needs 54–80°/s pan. Doc 3's stage 5 measures whether the motors stay silent at those speeds; if not, the coordinator caps indoor slew and lets the beam lag through close passes.
Hands don't work at range — design around them. The book story: anchor the person from feet/torso, offset to the wrist (body keypoints work at range), confirm "book" with open-vocabulary detection. Same evidence reshapes gestures: v1 gestures are arm-scale (raise, wave, hold-out — from body keypoints); finger gestures (pinch, palm shapes) are not feasible from a corner camera. Naive pointing rays err ~28° — fine for "which zone," not for precision (a trained pointing model reaches <2° if ever needed).
The failure modes are about feelings: never sweep eyes (predictive no-go cones + soft-edged dimming + a hard interlock — tracking loss parks the beam within a deadline); move legibly (a small anticipatory "glance" before the main move reads as intentional — HRI-validated); and suppress motion during movie/settled-reading — lights that move when people are trying to be still is the documented trust-killer.

2. User stories (the problem space)¶

ID	Story	Note
A — Follow & task
A1	Follow me across the room	The canonical hard story: tracking + smoothing + handoff
A2	Light the surface I'm using (table/board/desk)	Validated demand — $600 task lamps that don't even move
A3	Spotlight the book in my hands	White space: every current solution is worn on the body
A4	Cooking-step lighting (board → stove as activity shifts)	Context-driven; bridges to the DLP recipe story later
A5	Light my face on video calls	Adjacent shipped demand (ring lights, Apple Edge Light)
B — Ambient & idle
B1	Point at wall art when idle	Static target — trivial once calibrated
B2	Plant/decor accents on schedule	Zero live perception needed
B3	Gallery rotation with dwell times	Motion styling matters most here
C — Media & social
C1	Movie mode = stillness	Suppression is a feature; movement here is the trust-killer
C2	Light the board-game table when we gather	Group context from the activity model
C3	Party sweeps	The one home for stage-light behavior
D — Utility
D1	3am path light	Ankle-height, dim, never near eyes
D2	Find-my-keys by pointing light at them	No shipped product does this — needs object memory
D3	Closet/pantry assist	Beam into the space you just opened
D4	Pet play/follow	$20 auto laser toys validate the appetite
E — Interaction
E1	Arm-scale gestures (raise/wave/hold-out), context-armed	Gating evidence: 5.5× fewer false positives; V-JEPA-gated version is novel
E2	Point at a spot to send light there	Coarse zone-selection in v1; precision needs a trained model
E3	Summon / banish the beam	Beckon it over; wave it away
F — Multi-person & multi-fixture
F1	Each person gets their own beam	Research favors this "spatial multiplexing" over preference arbitration
F2	Seamless fixture handoffs while walking	Hysteresis so beams don't ping-pong
F3	Second fixture covers occlusions	Redundancy when line-of-sight breaks
F4	My preferences follow me	Identity via trajectory re-ID; feeds the roadmap's identity layer
G — Trust (gates everything)
G1	Never sweep a beam across eyes	Hard constraint: predictive no-go cones
G2	Fail calm	Tracking loss → park/dim, never hunt
G3	Legible motion	Anticipation micro-moves, context speed caps
G4	Visible "perception off" state	The being-watched objection is the adoption risk

3. The architecture in one picture¶

 PoE cams ──► RTX 6000 box:  go2rtc (low-latency restream)
                              ├─ FAST 15–30 Hz: detect (YOLO26) → track IDs →
                              │   ground-plane 3D → keypoints → One-Euro+prediction
                              │   → TARGET RESOLVER → ASSIGNMENT → SAFETY LAYER
                              └─ SLOW (event clips, 2–8 fps): V-JEPA → activity state
             Frigate: recording/review only — NOT in the control path
                              │ direct MQTT {fixture, pan, tilt, rate, dim}
             ESPHome C6 nodes (batch_delay: 0) ──► CAN ──► RMD actuators
             Home Assistant: observes, overrides — never in the loop

Annotations on each block:

Target resolver: static targets (art, table) are label lookups in the room model — zero runtime perception. Dynamic targets come from tracks, led by velocity. Book = wrist offset from the person anchor.
Assignment (which fixture takes which target): bipartite matching (Hungarian) with costs = incidence-angle badness (light the task from beside/behind the person, not their face) + travel + a hysteresis bonus for the incumbent. Run assignment at 1–5 Hz (sticky), aiming at 15–30 Hz — per-frame reassignment invites beam flapping. Handoffs cross-fade. Honest status: no published algorithm exists for lights; this is greenfield.
Safety layer (runs last, overrides all): predictive no-go cones around heads (extrapolate by angular velocity — static per-frame checks demonstrably fail on fast motion), soft-edge dimming at cone borders, UGR ≤ 19 as a computable glare budget, and the interlock: tracking-confidence loss → park/dim within a deadline.
The camera sees the beam — design for it: a bright moving spot shifts auto-exposure and casts moving shadows. Lock camera exposure, test the tracker with beams active, mask the commanded spot region if needed. (The same fact becomes an asset in calibration.)
Aiming math: closed-form angles per fixture from the calibrated model, with the dual-solution constraint (every target has a mirrored pan+180° twin — pick one and don't flip).

4. Step-by-step execution plan¶

Phase 0 — "The lamp follows me" (a weekend; needs only the Doc 3/4 rig + any RTSP camera)¶

Mount the camera high in a corner and get its stream working — budget an hour if you've never set up an IP camera: on first boot it forces an admin password (via its app or web page at the camera's IP); enable RTSP/ONVIF in its settings (often off by default); then confirm the stream plays in VLC (Media → Open Network Stream). Amcrest/Dahua URL pattern: rtsp://user:pass@CAMERA_IP:554/cam/realmonitor?channel=1&subtype=0 — the camera's web UI or manual confirms the exact path. While in settings, set fixed exposure (the beam-in-view rule from §3).

🤖 Give this to your agent

You're my bench agent for the Engineered Lighting perception stack
(chapter: engineering.engineered.lighting/05-teach-it-to-aim/, Phase 0
steps 2-6). The corner camera is mounted and streaming RTSP with fixed
exposure, and the Doc 3/4 gimbal rig is on the bench with its ESPHome
node subscribed to the spotlight/target MQTT topic, forwarding to CAN.

Start by proposing a plan and wait for my approval before executing
anything.

Task: build the follow-me tracker on the GPU box, in a git repo, and
iterate on it while I watch the beam. (a) pip install ultralytics and
run YOLO26 person detection on the stream with ByteTrack IDs via
model.track(); read frames through OpenCV with buffering disabled —
grab-and-discard so we always process the newest frame. (b) Floor
calibration: I tape 4 marks in a big rectangle and tape-measure their
true room-frame positions; take my clicked pixel positions from one
saved frame and compute the pixel-to-floor homography with
cv2.getPerspectiveTransform. (c) For each tracked person, map the
bottom-center of the box (the feet) through the homography to (x, y)
on the floor; smooth with a One Euro filter, estimate velocity, and
compute the led aim point (position + velocity × measured loop
latency). (d) Fixture pose: I tape-measure the gimbal's (x, y, z);
compute pan = atan2 and tilt = atan toward a chest-height point above
the aim point. (e) Publish {"v":1, "ts":..., "pan":..., "tilt":...,
"rate":60} — the Doc 6 envelope fields — to the spotlight/target
topic at ~10-15 Hz, and use mosquitto_sub on that topic to verify
what the fixture is actually being told; the 0xA4 speed field keeps
motion continuous between updates.

SAFETY — non-negotiable: never command motor motion unless I confirm
I'm watching with a hand near the supply switch. The bench current
limit stays as set. Announce each motion command before sending it
and wait for my explicit go.

Done when: the beam follows you smoothly across the room and you have
latency + noise numbers.

Report back: end-to-end latency (wave-to-beam from slow-mo video),
beam lag at walking speed, tracking noise while standing still, a
mosquitto_sub capture of spotlight/target, and the diff of every
script added to the repo.

How to run this prompt →

On the GPU box: pip install ultralytics + run YOLO26 person detection on the stream; add ByteTrack IDs (built into Ultralytics' model.track()); read frames via OpenCV with buffering disabled (grab-and-discard so you always process the newest frame — the classic seconds-of-lag trap). This script — and every GPU-box script in this doc — is ideal Claude Code territory: run it in a git repo on the box and have it scaffold, run, and iterate on the tracker while you watch the beam; it can also mosquitto_sub the output topic to verify what the fixture is actually being told.
Floor calibration (15 min): tape 4 marks on the floor in a big rectangle; measure their true positions with a tape measure (room frame = one corner of the room, x/y along walls); click their pixel positions in one saved frame; cv2.getPerspectiveTransform gives the pixel→floor homography.
For each tracked person: bottom-center of the box (≈ feet) → homography → (x, y) on the floor. Smooth with a One Euro filter; estimate velocity; compute the led aim point (position + velocity × measured loop latency).
Fixture pose: measure the gimbal's (x, y, z) with a tape measure; pan = atan2, tilt = atan toward a chest-height point above the aim point.
Publish {"v":1, "ts":…, "pan":…, "tilt":…, "rate":60} (envelope fields per Doc 6) to the MQTT topic spotlight/target at ~10–15 Hz (the name Doc 3 stage 9 and Doc 4's stage-6 look-ahead both use); the Doc 4 ESPHome node subscribed to that topic forwards to CAN. The 0xA4 speed field makes motion continuous between updates.
Walk the room. Measure: end-to-end latency (wave-to-beam on slow-mo video), beam lag at walking speed, noise during tracking.

Done when: the beam follows you smoothly across the room and you have latency + noise numbers. This single demo validates the entire stack's riskiest loop.

Phase 1 — Surfaces & suppression (the room model arrives)¶

🤖 Give this to your agent

You're my bench agent for the Engineered Lighting perception stack
(chapter: engineering.engineered.lighting/05-teach-it-to-aim/, Phase 1
— surfaces & suppression). The follow-me loop from Phase 0 works off
a 4-point floor homography, and I have a LiDAR room scan (Polycam
Space Mode, rescaled against one tape-measured distance) exported as
a mesh/point cloud.

Start by proposing a plan and wait for my approval before executing
anything.

Task: be my calibration helper for registering the camera to the room
scan and wiring in the room model. (a) Help me label ~10-20 surfaces
from the scan into named JSON entries {name, centroid, normal,
extent} — I do the lasso clicking in CloudCompare or SuperSplat, you
build and validate the JSON. (b) Register the camera to the scan: I
photograph the ChArUco board at a spot picked from the scan; you run
solvePnP and invert for the camera pose, with sanity checks —
reprojection under 1 px plus a few clicked non-coplanar points.
(c) Replace the Phase 0 4-point floor homography with the registered
camera + the scan's floor (better everywhere in the room). (d) Wire
activity context in: subscribe the states the existing model
publishes over MQTT and map them in the resolver — reading@couch to
wall-art idle, cooking to the cutting-board target, movie to
suppression (no spontaneous motion).

SAFETY — non-negotiable: never command motor motion unless I confirm
I'm watching with a hand near the supply switch. The bench current
limit stays as set. Announce each motion command before sending it
and wait for my explicit go.

Done when: "light the table," "point at the art," and movie-stillness
all work from label lookups.

Report back: the solvePnP reprojection error and non-coplanar check
results, before/after floor-position accuracy at a few tape-measured
spots, the labeled-surfaces JSON, a mosquitto_sub capture showing a
context state switching the target, and the diff.

How to run this prompt →

Scan the room (Polycam Space Mode, LiDAR); tape-measure one known distance; rescale if off. Export mesh/point cloud.
Label ~10–20 surfaces once, manually (CloudCompare or SuperSplat lasso → named JSON {name, centroid, normal, extent}). Fifteen minutes of clicking beats a research project — auto-labeling (open-vocab 3D) is still unreliable on loose language like "cutting board."
Register the camera to the scan: ChArUco board at a spot picked from the scan → one photo → solvePnP → camera pose (invert!). Sanity: reprojection <1 px, plus a few clicked non-coplanar points. (Marker-based beats automatic matching indoors — the leading auto tool's own maintainers report >80% indoor failure.)
Replace Phase 0's 4-point floor homography with the registered camera + scan floor (better everywhere in the room).
Wire activity context in: your existing model publishes states over MQTT; the resolver maps reading@couch → wall-art idle vs. cooking → cutting-board target; movie → suppression (no spontaneous motion).

Done when: "light the table," "point at the art," and movie-stillness all work from label lookups.

Phase 2 — Self-calibration (the crown jewel)¶

Greenfield — no agent prompt yet: no proven recipe to encode. Work this phase with your agent from the text below.

Sweep each fixture through a coarse pan/tilt grid spanning its full range but constrained to camera-visible landings (spots outside the FoV or occluded teach nothing; if the visible region is small, densify there or add a second camera pose).
At each grid point: beam on/off, frame-difference, centroid of the spot; ray-cast that pixel onto the scan mesh for the 3D landing point. Lock exposure; run at dimmed ambient — the system controls the room's other lights, so calibrate at night automatically for maximum contrast.
Fit fixture pose + encoder zero-offsets + axis non-orthogonality by least squares (Levenberg–Marquardt; the telescope-pointing term structure). Error budget: scan ~2 cm + registration ~2 cm + centroid ~2 cm → RSS ≈ 3.5 cm ≈ 0.67° at 3 m — inside the ~1° budget, not by miles; measure residuals rather than assuming them.
Verify: command the beam to 10 labeled targets, measure landing error with the camera. Keep the manual 4-point jog as bootstrap/fallback. Re-fit automatically if drift is detected (the camera checks every landing forever — the system notices its own miscalibration).

Done when: a fixture placed anywhere self-calibrates in ~2 minutes and lands within ~1° on demand. This is also when fixture #2 arrives.

Phase 3 — Book & gestures v1¶

Greenfield — no agent prompt yet: no proven recipe to encode. Work this phase with your agent from the text below.

Book: person anchor + wrist keypoint offset → aim; open-vocab confirm ("book" via YOLOE/SAM-3-class model on a wrist-region crop) to arm the behavior.
Gestures: arm-scale only (raise/wave/hold-out from body keypoints), armed by context state; coarse point-to-zone (ray from shoulder→wrist, snapped to the nearest labeled surface).

Done when: the beam lands on a held book within one beam-width; a raised-arm gesture reliably toggles it; and a normal evening passes with zero false gesture triggers.

Phase 4 — Coordination & safety hardening¶

Greenfield — no agent prompt yet: no proven recipe to encode. Work this phase with your agent from the text below.

Two fixtures, two people: Hungarian assignment @ 1–5 Hz with hysteresis; cross-fade handoffs; predictive no-go cones + soft-edge dimming + tracking-loss interlock; motion styling (anticipation micro-move, context speed caps). Done when: two people walk crossing paths and the beams hand off without flapping or ever crossing a face.

Phase 5 — Precision & identity¶

Greenfield — no agent prompt yet: no proven recipe to encode. Work this phase with your agent from the text below.

Trained pointing model (<2°) only if coarse pointing earns it; geometric/trajectory re-ID → per-person preferences (feeds the roadmap's identity layer).

Done when: a pointed arm selects the intended zone ≥9 times in 10, and two household members each get their saved warmth/brightness without asking.

5. The depth question, answered directly¶

Tier	Hardware	Covers	When
0 — ground plane + room model	none new	follow-me, all static targets, couch-seated (model priors), book (wrist offset), heads for safety	Start here
1 — better software	none	multi-camera fusion (halves jitter), occlusion-robust tracking, object memory	as stories demand
2 — one stereo camera in a hard room	$359–579	true 3D for reclined/floor-sitting, heavy clutter, tight multi-person; sharper self-cal	only when Tier 0 measurably misses
✗ monocular depth networks	GPU only	—	skip — jittery, 18–40 cm @ 4 m, worse than free geometry

Fisheye note: a 180°+ lens covers big rooms from one corner but needs a fisheye calibration model + dewarp before any of the geometry above — prefer two normal cameras when a room demands more coverage.

6. Risk register (honest)¶

Novel, no precedent: the V-JEPA + real-time-tracker dual path and world-model-gated gestures are unpublished combinations — analogs de-risk the pattern; the integration is ours to prove.
Coordination policy is greenfield — expect tuning cycles on costs/hysteresis; no literature will save us.
Slow-path latency is seconds-to-tens-of-seconds (clips → V-JEPA): context-driven target switches lag activity changes; stories must tolerate it, and anything sub-second belongs on the fast path by design.
Hands at range stay unsolved (2026) — stories are designed to wrist-level precision, which the beam width forgives.
No public benchmarks for this GPU on Frigate/DeepStream/V-JEPA — benchmark early; margins on adjacent hardware suggest ample headroom.
Edge-migration debt (strategic): the prototype is deliberately central-GPU, but the commercial product runs on a home-hub NPU. Cheap to migrate: YOLO-class detection, tracking, all the ground-plane/calibration/aiming math (it's just algebra). Expensive to migrate: SAM-3-class open-vocab segmentation, DeepStream-specific pipelines (NVIDIA-locked), full-size V-JEPA. Prefer the portable primitive wherever two options tie; treat every DeepStream convenience as prototype-only.
Product-kit camera transport must be reconciled: if the commercial kit's per-room cameras stream continuous WiFi video, a 5–10 room home lands exactly in the documented WiFi-collapse regime. The edge hierarchy in the Engineered Lighting strategy roadmap (an internal doc outside this series) is the answer — tier-1 vision on-device, ship metadata/clips, not video — the prototype's stream-everything pattern is a lab convenience the product must not inherit. Wired/PoE is the alternative at installation-friction cost.
Camera brand watch item (checked July 2026 — re-check before any camera purchase): the FCC Covered-List expansion targets Dahua/Hikvision procurement — residential unaffected, but firmware/inventory longevity is worth watching for Dahua-lineage brands.
Patents for counsel: US 12,190,542 (beam self-calibration), US 11,956,873 (multi-fixture follow-spot control), Forma (motorized recessed fixtures).
Trust is the real adoption risk: the same research validating demand documents the "being watched" objection — suppression states, legible motion, and a visible perception-off affordance are requirements, not polish.

Doc 5 · Teach It to Aim — Spatial Intelligence for Coordinated Spotlights¶

Bill of Materials — per room + one-time, ~$120–320/room¶

Concepts (plain English)¶

1. What we're building (the six findings)¶

2. User stories (the problem space)¶

3. The architecture in one picture¶

4. Step-by-step execution plan¶

Phase 0 — "The lamp follows me" (a weekend; needs only the Doc 3/4 rig + any RTSP camera)¶

Phase 1 — Surfaces & suppression (the room model arrives)¶

Phase 2 — Self-calibration (the crown jewel)¶

Phase 3 — Book & gestures v1¶

Phase 4 — Coordination & safety hardening¶

Phase 5 — Precision & identity¶

5. The depth question, answered directly¶

6. Risk register (honest)¶

Further reading (most load-bearing)¶