Selective Robotic Follow-Up Pruning for Mechanically Pre-Pruned Vineyards

1. Agronomic and Labor Need

Dormant pruning sets a grapevine’s crop: the spurs retained on a cordon, and the buds on each, fix the vine’s balance between growth and fruit.

The United States has 883,000 bearing acres of grapes, worth $5.83 billion in 2025 (USDA NASS), and nearly every acre is pruned each winter. The University of California’s 2021 San Joaquin Valley North wine grape study budgets 33 hours of hand pruning per acre at a loaded $20.30 per hour. Silwal and colleagues report that seasonal crews, rushing under short-term hiring and low wages, prune vines inconsistently.

Mechanization automated the speed, not the decision. Pre-pruners cut to a geometric envelope, so a hand crew often follows them. In a published per-acre model, pre-pruning plus hand follow-up costs $459.25, of which 18 hours and $378 is the follow-up. On that study’s 20 pre-pruned vines, the machine left 40% of 268 canes with fewer buds than the pruning rule retains and 35% with more (Silwal et al., 2022).

That labor is increasingly imported. USDA ERS reports about 385,000 H-2A positions certified in fiscal 2024, against just over 48,000 in fiscal 2005, with labor at 40% of expenses for fruit and tree nut operations. The unmet need is a follow-up pruner that cuts to the grower’s rule, across varieties, below the crew’s cost once capital is counted.

2. State of the Art

Cane-pruning systems. The University of Canterbury system (Botterill, Green and colleagues) reconstructs vines in 3D, lets an AI system decide which canes to prune, and cuts with a six-axis arm. Archie Jnr (Williams, MacDonald and colleagues, 2024) pruned 71.1% of 311 canes across 25 vines in a commercial vineyard.

Follow-up spur pruning. Carnegie Mellon’s Robotics Institute and Cornell’s Lake Erie laboratory (Kantor, Bates, Silwal and colleagues) deployed a research prototype, Bumblebee, behind an OXBO VMech pre-pruner. On 20 Concord vines it pruned 87% of prunable canes at 213 seconds per vine. Its successor NIFA award (ended August 2026) proposed combining learning with classical approaches.

Servoing-first systems. Gebrayel and colleagues (2026, DOI: 10.1109/LRA.2026.3664530) pair model predictive planning with point-cloud visual servoing, tested outdoors in wind.

Learning-first control. The strongest learned-control evidence comes from tree fruit:

• At Oregon State, a learned vision controller with force control raised laboratory success from 46% to 77%.
• In the lab, Jain, Grimm and Lee’s visuomotor policy succeeded in 7 of 20 trials against 2 of 10 for a point-cloud planner, a difference these samples cannot distinguish (Fisher exact p = 0.68).

No field head-to-head exists in the literature reviewed.

Cordon-level automation. A LIDAR-guided tractor system reports 7.3 seconds per vine (Felicetti et al., 2024, DOI: 10.1002/rob.22453; vine count not stated in the abstract).

Where the field is. Follow-up spur pruning has worked once in the field: one row, one variety, one arm per side, with a simplified rule. What has not been demonstrated is attended cost including capital. Bumblebee’s cycle time at the tested 537 vines per acre implies 31.8 run-hours per acre, against the 18 its cost model assumed, and that model counted neither an attendant nor capital.

3. Foundational Research

4. Competitive Landscape

Non-selective machines. Pellenc’s VISIO pre-pruner and Precision Pruner, Westside Equipment’s VMECH Chariot and OXBO’s VMech head (about $30,000, per Silwal) are mature, but each cuts to a line or envelope. Pellenc and Westside disclose no pricing.

Selective machines. No selective pruner is confirmed on general sale in the United States. Robotic Perception lists one for pre-order, PeK Agroline markets robotic branch cutting with selectivity and US availability unverified, and Vision Robotics describes a prototype dependent on financing. None of the three discloses funding.

Funded academic work. Carnegie Mellon’s $1.0 million NIFA award (ended August 2026), and New Zealand’s MaaraTech programme, reported by Rural News Group as NZ$16.8 million over five years, led by the University of Auckland.

In the literature reviewed for this brief, no result reports attended multi-arm cost per acre on pre-pruned vines.

5. Total Addressable Market and Public Benefit

Definition. Annual United States labor spend on the selective hand follow-up cut after mechanical pre-pruning.

Bottom-up. 883,000 bearing acres (USDA NASS, 2025) × 18 follow-up hours per acre (Silwal et al.) × a loaded $20.30 (UC Davis, 2021) to $23.75 per hour (WSU TB106E, 2024) = $365.40 to $427.50 per acre, giving an upper envelope of $323 million to $377 million per year if all bearing acreage were pre-pruned and hand-finished. The envelope overstates fully mechanized acreage and misclassifies acreage still hand pruned without a pre-prune, and the 2021 wage understates current cost. Full hand pruning at 32 to 33 hours per acre would be $574 million to $692 million, about 10% to 12% of crop value.

Top-down cross-check. MarketsandMarkets puts global agricultural robots at $17.73 billion in 2025, reaching $56.26 billion by 2030 (26.0% CAGR, November 2025); the US follow-up envelope is about 2% of that.

Serviceable market. The serviceable market is defined by variety class, where the field evidence applies, and assumes every acre buys a hand follow-up.

• Initial: Washington’s 17,000 juice-grape acres plus New York’s 33,000 acres, a follow-up envelope of $18 million to $21 million per year. This is flagged as under $100 million, and NASS publishes no type split for New York.
• Expansion: vinifera wine acreage in California (510,000), Washington (50,000) and Oregon (32,000), a follow-up envelope of $216 million to $253 million, once follow-up spur pruning is proven on vinifera cordons.

Payment pathway. No CPT or HCPCS code applies; growers pay by capital purchase (Internal Revenue Code section 179 expensing up to $2,500,000 in 2025) or per-acre service priced below the $378 hand follow-up.

Public benefit. Consistent pruning on a $5.83 billion crop, and an open benchmark the field can share.

6. Research Gap and HHA Contribution

What has not been done

Follow-up robotic pruning has worked once in the field. Attended cost including capital that beats the crew has not been shown.

Why the robot is attended. Silwal’s robot follow-up costs $80.64 per acre, or $4.48 per run-hour, and counts no attendant and no capital. Dormant vineyards host tying and brush crews, so employee exposure is predictable, and California’s Title 8 section 3441(b) requires an operator stationed at the controls of self-propelled equipment.

Breakeven. Pricing the attendant at the paper’s $21 per hour gives:

Breakeven run-hours per acre = $378 ÷ ($21 ÷ robots per attendant + $4.48 + capital per run-hour).

The capital term rests on stated assumptions:

• the paper’s $115,000 prototype estimate, amortized over five seasons with no financing;
• 1,220 run-hours per unit per season, 10 hours a day over about 122 dormant-season days;
• together, about $18.85 per run-hour.

Under those assumptions:

• One robot per attendant breaks even at 8.5 run-hours per acre, 14.8 before capital.
• Three robots per attendant break even at 12.5.
• At half the unit cost, three per attendant break even at 18.1.

The tested block implies 31.8, a required gain of 1.8 to 3.7 times.

Supervision ratio is rule-bound. Under California’s Title 8 section 3441(b) as written, each self-propelled platform carries its own attendant, so the one-robot case applies there (8.5 run-hours per acre, 3.7 times). Higher ratios apply where rules permit, under a variance, or with several arm modules on one operator-carrying vehicle, capital counted per vehicle. At 480 season run-hours, three per attendant would need 6.4.

Three coupled targets. Supervision ratio, run-hours per acre and manufactured unit cost jointly decide viability. In the literature reviewed, no published system has measured the first or attempted the third.

The gaps behind them:

• Intervention rate. In the orchard trial, cuts on 96 of 115 false detections were withheld by hand.
• Arms per side. Two arms working both sides at once are bounded by the 137-second one-side time, about 20.4 run-hours per acre.
• Variety coverage. Follow-up spur pruning has been measured only on Concord in the literature reviewed.
• Correctness and outcome. Professional-label correctness and next-season yield are unreported.
• Manipulation. No field head-to-head of learned and classical manipulation exists.

Why the originating groups have not closed it

The leading programs pursued full autonomy on episodic grants (2015 to 2026), with no cost-down path for a $115,000 prototype, and equipment makers sell acres per hour, which a slow selective cutter does not fit.

What HHA’s program would do differently

• Supervision and throughput. Two arms per side on a stop-and-go platform, with scan-once reconstruction. Cuts are assigned across arms per vine segment by a mixed-integer linear program minimizing dwell time. Every attendant intervention is logged by cause, and supervision ratio is projected from intervention time.
• Perception and coverage. Spur-versus-cane detection is trained on labels from target vineyards and gated on precision. Coverage spans Concord and at least one vinifera cordon system, with professional pruner labels as ground truth.
• Simulation. A procedural vine simulator built in MuJoCo from scanned vine architectures, with randomized lighting, cane geometry, cordon wire position and camera pose, pre-trains and stress-tests both manipulation arms before each dormant season.
• Manipulation, benchmarked in the field. RRT-Connect with position-based visual servoing, against a Proximal Policy Optimization policy:
  • Observations: 30 Hz wrist optical flow, end-effector pose, and a 500 Hz force-torque wrench filtered to 50 Hz.
  • Actions: six-degree-of-freedom velocity at 10 Hz, bounded at 0.1 m/s.
  • Reward: 100 per metre of progress, plus 0.5 times the cosine of perpendicularity, minus 10 per collision, minus 0.05 per step, plus 50 for a cut completed under 15 N.
• Outcome. A randomized complete block trial of hand follow-up, robot follow-up and pre-prune only, measuring retained-bud variance, cluster count and yield.

Team mapping. Benchmark, evaluation and scheduling map to Haedar Hadi. Trial design, detection and intervention analysis map to Hass Dhia. Cutter life, tolerances and unit cost map to Ahmed. This is integration, supervision and manufacturing engineering, not unsolved science.

7. Comparable Funded Projects

Carnegie Mellon University, USDA NIFA AFRI, award 2021-67021-35974, $1,000,000, September 2021 to August 2026. Autonomous grapevine pruning, the direct predecessor; its published outputs that we located address perception rather than attended cost.

Oregon State University, USDA NIFA AFRI, award 2020-67021-31958, $374,704, July 2020 to July 2024. Autonomous dormant pruning of fruit trees (acknowledged in You et al., 2022); AFRI funds pruning robotics at single-institution scale.

Cornell University, USDA SCRI, award 2026-51181-46280, $5,835,164, July 2026 to June 2030. A multi-purpose tree fruit robot for harvesting and thinning, citing SCRI’s mechanization priority.

Washington State University, PI Ming Luo, NSF Cyber-Physical Systems, award 2312125, $1,199,998, 2023 to 2027. Human-robot apple harvesting; an attended pruner is the same human-in-the-loop class.

University of Kentucky, PI Biyun Xie, NSF CPS, award 2437812, $1,179,584, 2025 to 2028 (robotic tomato phenotyping in large-scale greenhouses), and Michigan State University, PI Zhaojian Li, NSF with USDA NIFA, award 2540586, $400,000, 2026 to 2029 (occlusion-resilient apple harvesting).

USDA award records do not publish principal investigators.

8. Opportunity Assessment

TRL evidence chain

• Follow-up spur pruning, one arm per side: TRL 5 (20 Concord vines, Silwal et al.).
• Selective cane pruning: TRL 5 (25 vines, Williams et al.).
• Learned manipulation: TRL 4 (laboratory).
• Attended multi-arm operation at a capital-inclusive breakeven: TRL 3 (unmeasured).
• Vinifera follow-up spur pruning against professional labels: TRL 3.

The composite TRL is 4.

Technical risks as research questions

1. Can intervention time fall below one third of run time, supporting three units per attendant where rules permit, or do detection errors and reach near wire keep the ratio at one?
2. Can two arms per side reach 12.5 run-hours per acre on Concord, and can reconstruction, decision and cut fit a 10-second per-cut latency budget on a stop-and-go platform?
3. Does a per-block rule layer reproduce professional decisions at 90% or better on two trellis systems, and does learned manipulation beat classical manipulation in the field by 5 points or more at equal or lower time?
4. Displacement. Orchard Robotics ($22 million Series A, crop imaging) could add cutting, or Robotic Perception could reach US sale first. Mitigation: publish the benchmark, intervention taxonomy and breakeven model openly, making HHA’s contribution the evaluation standard and manufacturing path a fast follower would use.

Regulatory pathway

No FDA clearance applies.

• California, Title 8 section 3441(b). An operator must be stationed at the vehicular controls, which may be at a location on the vehicle with starting, accelerating, decelerating and stopping controls. Steering controls are also required where the machine steers other than by ground or furrow or exceeds 2 mph. The attendant therefore rides each platform: one attendant per platform under the rule as written.
• Cal/OSHA memorandum, August 30, 2024. It finds no violation where there are no employees and no employee access. It rescinded Cal/OSHA’s opposition to autonomous agricultural vehicles and proposed an advisory committee on vehicles under 500 lb, 20 horsepower and 2 mph, a class that bears on a future unattended configuration.
• Standards Board and precedent. The Standards Board denied petition 596 for driver-optional tractors in 2022. Cal/OSHA’s experimental variance to Monarch Tractor (August 6, 2021) is the regulatory precedent, and John Deere’s autonomous 8R (January 2022) is the product precedent.
• Other states. Washington and New York operator rules were not researched for this brief.
• Machinery safety. ISO 18497-1 to 18497-4 (2024).
• Europe. Regulation (EU) 2023/1230 applies from January 20, 2027, with third-party assessment where learning AI modules ensure safety functions. Deterministic safety functions avoid that trigger.
• Model updates. Models are frozen per dormant season and retrained off-machine. FDA’s Predetermined Change Control Plan guidance is not jurisdictional and serves only as a model for a pre-specified update envelope.

Regulation as a time moat, with limits. An attended safety case, ISO 18497-4 verification history and proprietary field data accumulate each dormant season. Southern-hemisphere and multi-site trials can narrow that clock, and the benchmark is open by design, so the advantage is a head start measured in seasons.

Proposed experimental approach, first six months

On an October start, months 1 to 2 build the simulator, classical baseline and labeling protocol with a professional pruner. Months 3 to 6 (dormant season 1) scan and label 300 or more pre-pruned vines on Concord and one vinifera system and field-test the baseline on at least 100 vines, logging every intervention. Month 6 gate: decision correctness of at least 85% against professional labels; below 75%, re-scope to operator decision support.

9. Team Fit

Hass Dhia, Principal Investigator. MS Biomedical Sciences (Wayne State University School of Medicine), medical school background including service as an anatomy teaching assistant, and an AI infrastructure architect with a physical-sciences foundation. Components: randomized complete block trial design and statistics; stereo reconstruction and spur-versus-cane detection; the intervention taxonomy and supervision-ratio analysis; the physics of cutting woody tissue near rigid wire.

Haedar Hadi, Co-Investigator. MS Computer Science, Boston University (Information Systems focus), with cloud and database architecture experience. Components: the Proximal Policy Optimization policy and MuJoCo simulator; the field head-to-head against classical manipulation; the mixed-integer cut scheduler; the open vine benchmark and its infrastructure.

Ahmed, Director of Manufacturing. Components: design for manufacturability of the arms, end effector and row-straddling platform against the unit cost the capital-inclusive breakeven allows; cutter blade life from field cuts; a tolerance budget for cut clearance near wire; a quality system and field-serviceable modular design.

The lab-to-production bridge. Most research proposals end at “it works in the lab.” This proposal includes explicit DFM milestones at every phase, ensuring that prototype decisions consider production scaling, tolerance analysis, and quality systems from day one. This addresses the valley of death between TRL 4-5 prototypes and TRL 7+ deployable systems, the gap where most funded research stalls. Here it is concrete: halving the $115,000 unit cost at three robots per attendant relaxes breakeven from 12.5 to 18.1 run-hours per acre.

What this team does not have, and how funding addresses it. HHA has neither viticulture expertise nor vineyard access. The budget funds an extension viticulturist subcontract to own pruning rules and outcome measures, grower cooperator agreements for dormant-season sites, and one field robotics engineer. The team supplies the supervision, coverage, evaluation and manufacturing layers prior programs did not.

10. Recommended Next Steps

Target funder programmes

• USDA NIFA AFRI, Engineering for Agricultural Production and Processing. AFRI funded both pruning-specific awards in Section 7; the best single-institution fit.
• USDA Specialty Crop Research Initiative. Legislative focus includes improved mechanization; HHA’s route is a coordinated project with a viticulture program.
• NSF Cyber-Physical Systems. An attended multi-unit pruner is a human-in-the-loop cyber-physical system.

Funding range. Single-institution comparables run $375,000 to $1.2 million; the program is scoped at $1.15 million over 24 months.

Twenty-four month milestone plan

Months count from an October start, so months 3 to 6 and 15 to 18 are dormant seasons.

• M1–2 R&D: MuJoCo simulator, classical baseline, labeling protocol and intervention taxonomy.
• M1–3 Manufacturing (Ahmed), from month 1: DFM baseline for arms, end effector and platform. Tolerance budget against cut clearance near wire. Unit-cost model tied to the capital-inclusive breakeven.
• M1–6 Regulatory (parallel track): Ride-on attended configuration with starting, accelerating, decelerating, stopping and steering controls, per Title 8 section 3441(b). ISO 18497-2 obstacle protection design review.
• M3–6 R&D, dormant season 1: 300 pre-pruned vines scanned and labeled on two trellis systems. Baseline field test on 100 vines with every intervention logged. Gate: decision correctness at least 85%; below 75%, re-scope to decision support.
• M7–12 R&D: Detection retrained on season 1 labels. Learned policy trained. Open benchmark released. Regulatory: ISO 18497-4 verification plan and season-locked release procedure.
• M7–14 R&D: Four-arm scheduler on a laboratory row. Gate: projected run time below 12.5 run-hours per acre at Concord density (three units per attendant; 8.5 at one), with per-cut latency under 10 s. Manufacturing (Ahmed): Season 1 cutter wear characterized and second-iteration end effector built. Gate: 10,000 bench cuts between blade changes.
• M15–18 R&D, dormant season 2: Randomized complete block trial with learned and classical manipulation head to head. Gates: below 12.5 run-hours per acre at three units per attendant, or 8.5 at one, per the applicable operator rule; intervention time under one third of run time; decision correctness at least 90%; retained-bud standard deviation no worse than hand. Learning is adopted only at a 5 point advantage or more.
• M15–20 Manufacturing (Ahmed): Field-serviceable modular platform. Gate: the bill-of-materials unit cost, amortized over 1,220 run-hours a season for five seasons at the measured run-hours per acre and supervision ratio, gives a cost per acre below $378.
• M19–24 R&D and program: Cluster counts and harvest yield. EU 2023/1230 applicability assessed. TRL 5 at product scope declared, or the gap documented. First-revenue milestone: a per-acre follow-up pruning service agreement with a grower cooperator for dormant season 3, priced below $378 per acre.