# Field Brief: AR Glasses for Cherry Ripeness Decisions

**Owner context:** James Grunsky — cherry orchards, Lodi / Stockton CA (incl. 38-acre trellis at Handel Road)  
**Lane:** FIELD — ripeness cues for California sweet cherries, what glasses must show in <~1 s, phone → wearable build path  
**Not in scope (this brief):** SO-ARM101 robot arm / sim-to-real (later handoff only)  
**Date:** 2026-09-04  
**Status:** First-cut product/tech brief for grower-operator review

---

## 1. Problem (one paragraph)

California sweet cherries are hand-picked at the firm-mature stage with stems (pedicels) intact; San Joaquin County near Lodi is the state’s largest producing county, and harvest windows are short (typically mid-April through early/mid-June statewide, with Lodi u-pick/commercial windows often mid-May into June). Pickers must decide “ready / wait / leave for next pass” fruit-by-fruit under canopy shade, bright sun, dust, and heat, while crews turn over seasonally. Wrong calls hurt packout and price: too early → under-color, low sugar, immature rejects; too late → soft fruit, bruising, shorter shelf life after hydrocooling and color/size sort at the packer. Labor is scarce and expensive; every second of hesitation and every re-pick pass costs money. AR glasses (or a phone prototype that proves the decision model) aim to put a consistent, near-instant readiness cue in the picker’s view so decisions match packer/quality standards without slowing hands or requiring each worker to carry years of Bing/Coral/Rainier judgment.

---

## 2. Ripeness decision model for pickers

### 2.1 What commercial maturity actually uses (CA)

Per UC Davis Postharvest and California industry practice:

| Signal | Role in decision | Typical CA / literature anchors |
| --- | --- | --- |
| **Skin color** | Primary field cue; most used for harvest timing | Minimum maturity: entire surface at least **light red** and/or **14–16% SSC** depending on variety. **Red mahogany** recommended for Brooks, Garnet, Ruby, Tulare, King. Dark sweet types (e.g. Bing) are often targeted toward **mahogany / dark mahogany** for quality markets. CTIFL-style 1–7 color scales are widely referenced in research. |
| **Soluble solids (Brix / SSC)** | Quality & consumer acceptance; legal maturity companion | CA minimum band often cited as **14–16%** by variety; consumer research on Brooks/Bing has argued for fuller color plus **≥~16% SSC** for better acceptance. Measured with refractometer (destructive) or NIR (instrument). |
| **Firmness** | Handling, packout, shelf life | Softens as color darkens; still critical for bruising risk. Measured with penetrometer / FirmTech-type tools — **not** visible to a camera as a direct number. |
| **Size / weight** | Price row-size; secondary maturity correlate | Correlates with SSC early; canopy position matters (interior fruit often less mature). |
| **Defects** | Grade / cull | Cracks, bird pecks, doubles/spurs, shrivel, decay, misshapen fruit. |
| **Stem condition** | Freshness signal at market | Green fleshy stems preferred; harvest leaves pedicel intact. |
| **Taste (SSC/TA)** | Final quality | Related to SSC, titratable acidity, and SSC/TA ratio — needs bite or lab. |

**Lodi / San Joaquin variety context (relevant neighbors & local plantings):** Bing (traditional Lodi/San Joaquin dark sweet), Coral Champagne / Coral, Brooks, Rainier (blush/yellow), Lapin, Tulare, and related dark sweets. Exact mix on Handel Road is for James to confirm — models must be **variety-aware** (Rainier color ≠ Bing color).

**Important physiology note (UC Davis):** Cherry response to ethylene is minimal; ethylene does **not** accelerate ripening. You cannot “pick green and ripen like a banana.” Field color/SSC at pick largely locks quality trajectory.

### 2.2 Camera-visible vs needs touch / instrument / taste later

| Camera / vision can approximate | Needs touch, bite, or instrument |
| --- | --- |
| Skin color class (e.g. light red → red → mahogany → dark mahogany; Rainier blush %) | Exact °Brix |
| Rough fruit size / clustering | Firmness (g or FirmTech) |
| Occlusion / “can’t see enough — don’t guess” | SSC/TA taste balance |
| Visible defects (cracks, doubles, bird damage, obvious rot) | Internal browning / soft spots not yet showing |
| Canopy lighting context (sun vs deep shade) for confidence | Stem pull force / retention mechanics |

**Practical picker rule for MVP:** Treat **color class + visibility confidence** as the AR decision; use spot **Brix/firmness checks** (crew lead or QC) to calibrate the color thresholds by variety and block each morning — same as how good crews already spot-check.

### 2.3 False-positive / false-negative risks

| Error | What happens | Operational impact |
| --- | --- | --- |
| **False READY (too early)** | Under-color, low SSC, immature on grade | Packer rejection, brand risk, wasted labor into bins that sort out |
| **False WAIT when actually ready** | Fruit left; next pass softens or rains crack risk near harvest | Lost premium size/color window; more soft fruit later |
| **False READY (too late / soft)** | Soft fruit, bruising in lug/bin | Culls, short shelf life after hydrocool |
| **Rainier / light cultivars mis-modeled as dark** | Blush patterns misread as “not ready” or “ready” | Systematic bias on yellow/blush blocks |
| **Shade / backlight glare** | Color looks darker or washed | Inconsistent calls across row sides / times of day |
| **Partial view (leaf, neighbor fruit)** | Model guesses from incomplete surface | Overlay must show **MAYBE / NEED BETTER VIEW**, not READY |

**Design implication:** Prefer **conservative READY** (high precision on “pick now”) over high recall. Leaving a borderline cherry is cheaper than flooding the line with immature fruit.

---

## 3. AR UX requirements

### 3.1 What the overlay shows (keep it brutal and fast)

Minimum viable HUD states (color-blind-safe: shape + color + optional short word):

| State | Visual | Meaning |
| --- | --- | --- |
| **READY** | Green check / filled circle | High confidence: color class at pick threshold for this variety/block |
| **WAIT** | Amber pause / open circle | Immature by color; leave for later pass |
| **MAYBE** | Blue question / dashed | Borderline color **or** partial occlusion / bad light — picker uses judgment or repositions |
| **DEFECT** (optional MVP+) | Red X | Obvious crack/rot/double — don’t pick for pack |

Optional secondary line (small, peripheral): variety tag (`Bing` / `Coral` / `Rainier`) if crew is on mixed rows; confidence % only if it does not clutter.

**Do not** dump Brix estimates as fake precision unless calibrated NIR is in the loop. Do not animate 3D meshes over fruit in Phase A/B — latency and distraction cost more than value for picking cadence.

### 3.2 Latency budget

- **Target end-to-end:** detect → classify → overlay **< 1 second** from fruit entering clear view (stretch goal **< 300–500 ms** for natural picking rhythm).
- Research precedent: orchard sweet-cherry maturity detectors (e.g. improved YOLOX variants) report real-time-class inference on GPU; phone edge models can approach interactive rates if resolution and class count stay small.
- If inference > 1 s: **fail soft** — hide overlay rather than show stale READY on the wrong fruit.

### 3.3 FOV, occlusion, lighting

- Pickers work inches to ~arm’s length; FOV need is **narrow and stable**, not wide VR. Monocular assisted-reality HUDs (~15–20° display FOV) are enough for a status badge near the gaze target.
- **Occlusion is normal** in dense canopies / trellis: require enough visible surface (e.g. majority of fruit silhouette) or force MAYBE.
- **Sun / shade:** California orchard light swings hard. UX must include auto-exposure awareness and a **low-confidence** path; outdoor-readable displays (≥~2000 nits assisted reality; consumer display glasses claiming ~5000 nits still struggle at extreme glare).
- **Night:** not primary for fresh cherry harvest (day crews). Low priority.

### 3.4 Language / crew UX

- Many San Joaquin harvest crews are Spanish-primary. Ship **bilingual UI** from day one: icons first; words `LISTO` / `ESPERA` / `DUDOSO` (or equivalent James prefers) toggled with English.
- Prefer **voice-off by default** during pick (noise, distraction); optional voice for supervisors/scouts.

### 3.5 Distraction & safety

- Overlay must be **glanceable peripheral**, not a full-screen film over both eyes.
- No continuous video streaming to cloud required for the pick decision (privacy + bandwidth + battery).
- Hands stay free for ladders, lugs, stems — **no phone in hand** for Phase B.
- Dust, sweat, PPE, and heat: consumer fashion glasses are a training/data tool; industrial assisted-reality is the shift tool.

---

## 4. Hardware reality check (outdoor orchard)

Honest bands for dusty, hot, bright San Joaquin picking:

| Option | Price band (approx.) | Pros for orchard | Cons for picking |
| --- | --- | --- | --- |
| **Phone camera prototype** | Device already owned; mount/clip <$50–200 | Fastest MVP; best cameras; easy labeling app; no procurement fight | Occupies a hand or awkward chest mount; glare on screen; not “glasses” |
| **Meta Ray-Ban (camera AI)** | Consumer ~$300–500 class (Gen 2 claims up to ~8 h typical use) | Lightweight; good for **hands-free capture** / scouting; Meta Wearables toolkit emerging; Agerpoint-style ag pilots on Meta glasses | Limited or no true pick-status HUD on camera-only models; not industrial dust/IP; continuous AI drains fast (~tens of minutes reported for heavy Live AI) |
| **Meta Ray-Ban Display** | ~$799 | In-lens monocular HUD; reported ~5000 nits; small ~20° FOV suitable for badges | IPX4 (splash), not dust-proof industrial; heat/sweat; narrow FOV; consumer durability in trellis dust unknown |
| **RealWear Navigator 500** | ~$2,500–2,600 | **IP66**, outdoor-visible display, hot-swap battery, voice in noise, 48 MP camera, -20–50 °C, drop-rated — built for dirty outdoor work | Assisted reality (boom display), not immersive AR; cost; looks industrial (crew acceptance TBD) |
| **Vuzix (Blade 2 / M400 family)** | Industrial / mid–high | Sunlight-oriented nits on some models; used in field AR prototypes (e.g. vine pruning R&D on Blade 2 + Jetson); enterprise Android | Blade 2 battery often short (~2–2.5 h class in field writeups); weight/comfort; still pricey |
| **XREAL / similar tethered AR** | Prosumer | Nice displays indoors | Typically **weak for all-day outdoor rugged** picking; tether/phone dependency; dust/heat not primary design |

**Recommendation:** Prove the **decision model on phone** in James’s orchard first; for wearable pilot, prefer **assisted-reality industrial (RealWear-class)** or a **bright monocular display glass** only after phone metrics work. Do not buy a fleet of fashion AR until false-READY rate is known.

---

## 5. MVP build path (3 phases)

### Phase A — Phone camera prototype in orchard (this season or next)

| | |
| --- | --- |
| **Deliverable** | Android/iOS app: point at cluster → READY / WAIT / MAYBE in <1 s; variety selector; offline-capable model; simple logging (GPS/time/block optional). Companion “label mode” for QC. |
| **Success metric** | On held-out Lodi images + live spot checks: **precision on READY ≥ ~90%** vs crew-lead/packer color standard; median latency <1 s; does not increase pick time vs unaided on a timed row trial (or <5% slowdown). |
| **Effort (order of magnitude)** | **4–10 person-weeks** software + **2–5 orchard half-days** for data/calibration (one engineer + James/QC). Hardware cost: phones on hand. |
| **Blockers** | Access to blocks during color change; variety list; agreement on color chart (UC Davis / CTIFL / packer card); crew time for labeled spot samples with Brix subset. |

### Phase B — Ruggedized glasses or assisted-reality HUD

| | |
| --- | --- |
| **Deliverable** | Same model on RealWear-class **or** Display glasses: hands-free badge overlay; bilingual; shift battery plan; dust wipe procedure. |
| **Success metric** | Full-shift usability (comfort, battery swaps); READY precision holds vs Phase A; **≥1 crew** prefers or matches unaided quality with equal or better bins/hour on pilot rows. |
| **Effort (order of magnitude)** | **8–20 person-weeks** after Phase A (device integration, UX, durability). CapEx: ~$800–$3k per pilot unit depending on platform. |
| **Blockers** | Outdoor readability under Lodi sun; sweat/dust failures; crew acceptance; bilingual voice vs silent HUD; procurement/support. |

### Phase C — Path toward robot vision handoff (later)

| | |
| --- | --- |
| **Deliverable** | Shared **labeled dataset + color/ripeness ontology** and export of detections (bounding boxes, class, confidence, variety) suitable for SO-ARM101 / orchard robot perception — not the arm itself. |
| **Success metric** | Dataset card + model checkpoint reusable in sim; documented domain gap (human egocentric vs fixed/robot camera). |
| **Effort (order of magnitude)** | **Ongoing** data ops; **2–6 person-weeks** packaging once Phase A data exists. |
| **Blockers** | Different cameras/viewpoints; real-time control latency ≪ picker HUD; safety around humans — out of FIELD lane. |

---

## 6. Data needed (without slowing harvest)

### 6.1 What to collect

- **Egocentric / phone** RGB images and short bursts of **Bing, Coral, Brooks, Rainier** (and any Handel Road cultivars) under **morning / midday / late**, **sun / shade**, **trellis angles**.
- Labels: color stage aligned to a **written chart** (e.g. light red / red / mahogany / dark mahogany; Rainier blush bands); binary **pick decision**; optional **defect** tags.
- **Calibration subset (small):** same fruit → photo → refractometer Brix + firmness note — enough to map color→quality per variety/block (tens to low hundreds of fruit per major variety, not thousands).
- Negatives: leaves, doubles, cracked fruit, distant fruit, motion blur.

### 6.2 How to collect without killing harvest pace

1. **Scout pass, not pick pass:** one trained person walks ahead or uses end-of-row 10-minute blocks with phone mount; pickers uninterrupted.
2. **QC lug sample:** at hydrocool/grade, photograph sorted color classes already decided by packer — free “labels” from operational truth.
3. **Morning calibration kit:** 20 fruit/variety → photo + Brix; update app thresholds; then put phones away.
4. **Wearable capture later:** Meta-class glasses good for **hands-free video dumps** after shifts for labeling — not required for Phase A.
5. Prefer **on-device storage**; sync Wi-Fi at shop. Avoid cellular upload of continuous video.

**Target scale for a useful first model (order of magnitude):** low thousands of labeled fruit instances across lighting conditions (research cherry maturity sets have used on the order of ~2k orchard images / tens of thousands of instances — treat that as an upper aspiration, not a week-1 requirement). Start with **hundreds of well-labeled instances per major variety**, then expand.

---

## 7. Open questions for James (max 5)

1. **Which cultivars and rootstock/trellis blocks** on Handel Road (and other ranches) should Phase A prioritize in 2027 harvest — Bing only, or Bing + Coral + Rainier in year one?
2. **What is the packer’s / buyer’s color card** you actually grade to (UC Davis chart, CTIFL, Stemilt/in-house), and is the commercial target **minimum legal maturity** or **premium mahogany + high Brix**?
3. **Crew language mix** and willingness: Spanish-first silent HUD OK? Any ban on cameras/glasses in orchard for privacy or labor-relations reasons?
4. **Who owns morning Brix/firmness spot checks** today, and can they add a 15-minute photo+Brix calibration without adding a headcount?
5. **Pilot constraint:** Is the goal “help inexperienced pickers match veterans” or “raise veteran consistency / reduce second picks” — that choice sets READY precision vs recall?

---

## 8. Sources (real URLs from research; no fabricated papers)

### Ripeness / CA cherries
- UC Davis Postharvest — Cherry produce facts (color + SSC maturity): https://postharvest.ucdavis.edu/produce-facts-sheets/cherry  
- California Cherries — industry facts (hand harvest, color, San Joaquin/Lodi, hydrocool/sort): https://calcherry.com/facts/  
- California Cherry Board research review PDF (2026; color scale / quality sampling context): https://calcherry.com/wp-content/uploads/2026/02/Cherry-Research-Review-2026.pdf  
- Harvest maturity indices overview (color, firmness, SSC; CTIFL scale mentions): https://cherrytimes.it/en/news/sweet-cherry-harvest-maturity-indices-turkey-usa-afghanistan-denmark  
- Crisosto et al. consumer quality / color–SSC discussion (Brooks/Bing; CA code vs proposed higher index): https://ucanr.edu/?legacy-file=162053.pdf&legacy-file-path=sites%2Fbtfnp%2Ffiles%2F  
- Rainier maturity indicators (SSC–color correlation): https://doi.org/10.21273/hortsci.22.2.283  
- Harvest timing / sensory (Bing, Sweetheart context): https://doi.org/10.21273/horttech.19.4.748  
- CA CCR cherries standards (mature, defects): https://www.law.cornell.edu/regulations/california/3-CCR-1428.4  

### Local variety context (Lodi / Stockton area examples)
- Lodi Blooms / Chinchiolo — Bing, Brooks, Coral Champagne, Lapin, Tulare: https://www.chinchiolofarming.com/pages/lodi-blooms  
- John Podesta Farms — Coral, Bing, Brooks, Rainier: https://www.johnpodestacherries.com/  
- Chinchiolo Stemilt — Bing, Coral, Rainier (Stockton): https://ccof.org/directory-member/chinchiolo-stemilt-california/  

### CV / AR / smart glasses in fruit & orchards
- Gai et al., *Agronomy* — real-time sweet cherry maturity detection (YOLOX variant) in natural orchard: https://doi.org/10.3390/agronomy12102482  
- *Computers and Electronics in Agriculture* (2025) — orchard sweet cherry color distribution from video + sensors: https://doi.org/10.1016/j.compag.2025.110334  
- 3D2cut AI/AR pruning glasses (Vuzix Blade 2 + Jetson; R&D): https://www.3d2cut.com/ai-ar-pruning-glasses/  
- Agerpoint + Meta AI Glasses Catalyst Grant; CA permanent-crop pilots: https://www.agrospectrumasia.com/news/26/4389/agerpoint-secures-meta-catalyst-grant-to-bring-ai-powered-smart-glasses-to-agricultural-field-operations.html  
- Meta AI Glasses Impact Grants (includes ag wearable examples): https://about.fb.com/news/2026/01/ai-glasses-impact-grants-wearable-technology-for-good/  

### Hardware
- RealWear Navigator 500 retailer/specs (~$2,600; IP66; outdoor display): https://www.exloc.com/products/realwear-navigator-500-rugged-handsfree-tablet  
- RealWear Navigator 500 user guide (IP66, outdoor display, hot-swap battery): https://support.realwear.com/hubfs/Knowledge%20Center/RealWear%20Navigator%20500/RealWear%20Navigator%20500%20User%20Guide%20v1.2%20English.pdf  
- Meta Ray-Ban (Gen 2) battery/video: https://about.fb.com/news/2025/09/ray-ban-meta-gen-2-better-battery-life-video-capture/  
- Meta Ray-Ban Display product page (~$799): https://www.meta.com/ai-glasses/meta-ray-ban-display-glasses-and-neural-band/  
- Display glasses outdoor nits / FOV reporting: https://tech.yahoo.com/ar-vr/articles/everything-know-meta-ray-ban-163736660.html  

### Unknown / not verified here
- Exact cultivar list and packer color card on James’s Handel Road block — **ask James**.  
- Published peer-reviewed **AR glasses cherry-picking** field trials with packout ROI — **not found** in this pass; closest are pruning AR (vines) and Meta/Agerpoint scouting wearables, plus RGB maturity CV papers.

---

## 9. Immediate next step (actionable)

**Before or at first color-break next harvest:** run Phase A — one phone, one variety (likely Bing), one trellis row, UC Davis–aligned color labels + 20 fruit Brix morning check, timed READY/WAIT overlay vs veteran picker. If READY precision holds, schedule a 1-unit RealWear or Ray-Ban Display pilot for Phase B; package labels for robot vision later (Phase C).