Camera, Sonar, or Laser Grid

How the first build of the second-generation Third Eye Kit should sense the world. Every number below comes from a datasheet, a store listing, or a published study, linked in the references.

Revisions since publication. This study was the first piece of applied research in the second-generation program, and later work refined it. Design review revised the latency accounting: about 70 ms from the world to the skin for a hazard ahead and about 120 ms for the ground, modeled from the parts' datasheets until the bench measures it, with the forward laser grid in its 4 by 4 mode and the ground grid in 8 by 8. The placement study set the ankle bands at four motors, not three. Design review moved both laser grids into the waist sensor pod and made the rear channel wait for a body-worn radar walk test. The hardware panel corrected the sonar cone to under 50 degrees and the sensor stack's cost figures are estimates that live in HQ. Where a number below has been superseded, the revision note says so. See the design review record and Where the body should feel.

The verdict in three sentences

For the first build, it earns no place on the vest.

Plain ultrasonic sonar cannot replace it. One sonar returns one blurry number, it goes nearly deaf to people in soft winter clothing, and a belt of eight sonars needs about a quarter of a second per sweep because they must take turns.

No AI anywhere in the path.

Physics does the seeing. The skin does the reading. The phone, later, does the talking.

The question: how should the first build feel the street?

The reference spec froze a founding rule: life-critical sensing and steering is physics, not AI inference. The open question for the first build was which physics. The candidates are the four ways a machine can measure space: bouncing sound (ultrasonic sonar), bouncing invisible laser light (time-of-flight, ToF), bouncing radio waves (millimeter-wave radar), and interpreting pictures (a depth camera plus AI).

These four are not rivals for one seat. They fail in different places, which is exactly why the spec calls for fusion. The real decision is which one carries the main load in the first build, and which ones ride along as backup. That decision is what the numbers below settle.

The mechanism: what each sensor actually hands the processor

The whole argument turns on one thing: the shape of the data each sensor produces, and how much brain it takes to turn that data into a buzz on the skin.

The founder's instinct, checked: the camera really is the expensive road

The founder guessed a camera is expensive, power hungry, and needs a bigger brain. All three guesses are correct, and now they have numbers.

• Power. The camera draws 2.5 to 4.5 W and the Jetson 7 to 15 W under real load. Call it 10 to 20 W on the body. A full day at that draw means a laptop-class battery, roughly half a kilogram of it. The laser grid, radar, sonars, microcontrollers, and motors together run on 1 to 2 W: a pocket power bank.
• Brain. Pixels mean nothing until a neural network turns them into distances and labels. That forces a Linux computer, a boot time, a thermal budget, and a whole category of software failure onto a safety device. The spec's founding rule exists for exactly this reason: the fast layer must be physics.
• Speed. A vision pipeline typically lands in the 100 to 500 ms range end to end. The camera cannot live below the isolation boundary; the direct sensors can. (Revision: the engineering record, docs/14, budgets the controller's own share, readout and fusion, at up to 10 ms ahead and 16 ms for the ground; the world to skin figure for the whole chain is about 70 ms ahead and about 120 ms for the ground, modeled.)

The camera is postponed, not banished. Sign reading and scene description stay in the plan, in the assistant role the spec already gave them.

The correction: where plain sonar quietly fails

Here is the part of the original idea that the evidence pushes back on. But three physical facts cap what a sonar belt can ever become, and each one matters most for exactly this user.

• It sees a cone, not a picture. An HC-SR04 hears everything inside its cone as one echo (about 15 degrees each side at the half-power point; the hardware panel later measured the full cone at under 50 degrees). It can say "something within 2.4 m somewhere in this cone" and nothing more: no shape, no left versus right inside the cone, no "two things." An environment picture is precisely what one sonar cannot produce.
• Soft clothing swallows it. 40 kHz sound is absorbed by foam, heavy fabric, and winter coats, and curved surfaces like a human shoulder scatter the echo away from the receiver. Sensor guides state it plainly: for soft or curved targets, ultrasonic is the wrong technology. A pedestrian in a coat is a soft, curved target. The most important obstacle on any sidewalk is the one sonar is worst at.
• A belt of them is slow. Fired together, neighboring sonars hear each other's pings (crosstalk), so they must take turns, about 33 ms apart. Eight sonars per sweep is roughly 264 ms between updates of any one direction. The laser grid refreshes all 64 directions in 66 ms, and its narrow light beams do not interfere at all.

Sound does keep one superpower light does not have: glass. A glass door faced squarely is nearly invisible to a laser and to a camera, but it is a mirror for ultrasound. That is why sonar stays in the stack, demoted to two jobs it is genuinely best at: catching glass and glossy surfaces, and dirt-cheap close-range redundancy right where the ToF grid has its minimum-distance blind zone. (Revision: the record turns the two sonars about 20 degrees outward so an angled glass door returns at least one echo, docs/04; whether it does every time is a bench question of G0, docs/13.)

The full comparison: four technologies, one table

Verified figures at the time of the study. Prices are single-unit street prices.

Read the columns and the fusion writes itself.

Prior art: the market and the research both point the same way

Notice the empty seat. Nobody ships a low-cost, worn, multi-sensor system whose primary output channel is a body-wide buzz language. That empty seat is the same market gap the reference spec identified, now confirmed from the sensor side.

The design: the laser grid maps straight onto the buzz language

This is the most practical finding of the whole study, and the reason the choice is not close. The buzz language needs direction and distance. The 8 by 8 grid is direction and distance, already sorted into zones.

The build: two costed first-build stacks

Both stacks share one architecture, faithful to the first generation's multi-controller lesson: a dedicated reflex microcontroller (RP2040 class) runs only ranging and motor drive in a hard loop, and a second spine microcontroller (ESP32-S3) handles logging and Bluetooth to the phone, so nothing slow can ever stall the buzz. The figures are estimates at the time of the study; the current parts list lives in HQ and design review has since revised the stack (two laser grids in the waist pod, four motors per ankle band, an actuator bench set before the order).

Stack A. Frugal bench rig

Stack B. Standard first build, the recommended stack

For comparison: the camera route's entry ticket is about double Stack B's total, before it senses anything at close range, and it still needs the direct sensors anyway. Even Glide, a camera-first product, ships near-range sensors beside its camera.

The free brain: the phone already in the pocket

The camera question has a quiet, practical answer that costs nothing: the wearer's phone. In the first build, the ESP32 streams over Bluetooth to a phone, which gives three things on day one:

• Logging. Every frame, every buzz, every button press lands in a file. This is what makes the first build an instrument that can answer the spec's open questions about legibility and false positives.
• The assistant layer, already built. Scene description and sign reading exist today as free phone apps. The wearer points the phone and asks. That is exactly the assistant role the spec allows: on demand, never steering, never in the safety loop.
• The first generation's family features, upgraded. The GPS and GSM texting node from the first generation becomes a phone feature, with no extra hardware on the body.

A dedicated onboard computer stays on the roadmap for the day the kit must work without a phone. The first build does not need it to prove the thing the second generation exists to prove.

Honest limits: what this choice must still prove at the bench

The program's culture is to name the unproven parts, so here they are for the sensing stack. These became the bench checklist, later expanded in design review to eight tests including the body-worn radar walk:

• Direct summer sun. The laser grid's range shrinks in strong ambient infrared. The datasheet says it works outdoors; how far, on a bright day, on dark clothing, is a number to measure, not assume. The radar and sonar columns exist partly as the sunny-day safety net.
• The glass door test. Walk the rig at a glass storefront. The ToF should miss it; the sonar must catch it every time when faced squarely. If it does not, sonar placement changes.
• The winter coat test. A person in a heavy coat at 1, 2, and 3 meters. ToF and radar must both see them at every distance where sonar goes quiet.
• The curb test. The down-tilted grid must flag a 15 cm drop at walking pace with enough warning to stop, indoors and out.
• Motor count versus battery. Sixteen buzzing LRAs under real use is a power draw to measure, not estimate from datasheets.
• Crosstalk inside the stack. Two laser grids and two sonars firing near each other need an interleave scheme; the firmware must prove the frame survives it.

None of these are reasons to hesitate. They are the first week of bench work, and every one produces a number for the record.

Recommendation: what to do next

• Adopt the fusion ladder: laser grid ToF as the workhorse, mmWave radar for motion and weather, ultrasonic for glass and close redundancy, camera postponed to the phone.
• Order Stack B, plus one spare of each sensor for the bench (revised in design review: after milestones G0a and G0b).
• Draw the practice schematics around the two-microcontroller architecture, then hold the compute discussion with real wiring in hand.
• Write the buzz-language table (zones, rhythms, the drop-off and heat signatures) before firmware, since FIG. 02 shows the firmware is mostly that table.
• Run the bench tests and publish the numbers, whatever they say, in the same honest register as the reference spec.

References

• ST VL53L8CX product page and datasheet: 8 by 8 zones, 400 cm, 215 mW continuous, 15 Hz at 8 by 8, 60 Hz at 4 by 4. st.com, price at estore.st.com, mouser.com
• shillehtek.com; RD-03D: 8 m, 3 targets. openelab.io
• shop.luxonis.com
• Jetson Orin Nano power: 7 W and 15 W modes, 8 to 16 W under real inference. edgeaistack.ai
• HC-SR04 limits: beam angle, 33 ms crosstalk interleave, soft-material absorption, curved-surface scatter. industrialmonitordirect.com, envistiamall.com
• NOA by biped.ai: three RealSense D430 depth cameras, shoulder-worn computer. realsenseai.com, biped.ai user manual
• Glide by Glidance: stereo depth camera plus near-range sensors, wheeled. [newatlas.com](https://newatlas.com/good-thinking/glide-white-mobility aid-blind-assistive-device)
• ALVU, MIT CSAIL: ToF lidar belt and vibrotactile strap for safe local navigation. researchgate.net
• Head-mounted three-sensor VL53L1X ToF plus haptics study: detection validated 15 to 150 cm, haptics directionally accurate and distance-proportional. preprints.org
• Third Eye Kit second-generation reference spec and founding rules. teww.org, the dossier on this site
• Revisions: the design review record, Where the body should feel