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

1. Problem Statement

Every dormant season, United States grape growers pay people to make a decision that no machine confirmed on general sale makes: which spur on a cordon stays, and which goes.

USDA NASS counts 883,000 bearing acres of grapes in 2025, with a value of utilized production of $5.83 billion (Noncitrus Fruits and Nuts 2025 Summary, May 2026). By state:

State	Bearing acres	Type split
California	751,000	510,000 wine, 126,000 raisin, 115,000 table
Washington	67,000	50,000 wine, 17,000 juice
New York	33,000	no type split published
Oregon	32,000	no type split published

Hand pruning is expensive and well documented. The University of California's 2021 San Joaquin Valley North wine grape cost study budgets 33 hours of hand pruning per acre at a loaded $20.30 per hour, with no mechanical pre-prune in the budget. It notes that much of the region's pruning is paid by piecework (UC Davis and UCCE, 2021).

Growers who mechanize still pay for the decision. A published cost model built on New York minimum-wage assumptions prices full hand pruning at 32 hours and $672 per acre. It prices mechanical pre-pruning at $81.25 per acre, plus an 18-hour hand follow-up at $378, for $459.25 combined (Silwal et al., Field Robotics, 2022). The follow-up exists because pre-pruners cut to a geometric envelope: on that study's 20 test vines, the pre-prune left 40% of 268 canes with fewer buds than the pruning rule retains, and 35% with more.

At $20.30 to $23.75 per hour (the latter Washington State University's 2024 orchard rate), 18 hours per acre across all bearing acreage gives an upper envelope of $323 million to $377 million a year. That is the spend if every acre were pre-pruned and hand-finished. The figure errs in both directions:

• It overstates spend on acreage pruned fully mechanically.
• It misclassifies acreage still pruned entirely by hand, which spends more per acre but is not follow-up.

Full hand pruning at 32 to 33 hours per acre would put the gross envelope at $574 million to $692 million, about 10% to 12% of crop value.

The labor is getting harder to buy. USDA ERS reports:

• About 385,000 H-2A positions certified in fiscal 2024, against just over 48,000 in fiscal 2005.
• A 2024 farm wage of $18.12, against $30.13 nonfarm.
• Labor at 40% of expenses for fruit and tree nut operations.

2. State of the Art

Non-selective pre-pruners are mature. All of them cut to a line or envelope, and none decides spur by spur.

Machine	What it does	Stated speed or claim
Pellenc VISIO	Pre-prunes, with vision-based automatic opening around posts	3.1 mph
Pellenc Precision Pruner	Tracks the cordon by vision	Up to 2 mph; maker claims up to 90% less pruning time
Westside Equipment VMECH Chariot	Pre-prunes and near-finish prunes	Up to 3 acres per hour; maker claims savings up to $600 per acre (baseline not stated)
OXBO VMech head	Pre-prunes; used in the Silwal study	Not stated

A LIDAR-guided tractor system reports cordon pruning at 7.3 seconds per vine (Felicetti et al., Journal of Field Robotics, 2024, DOI: 10.1002/rob.22453; the abstract does not state the number of vines tested).

Selective robots have worked in commercial vineyards, in research trials.

• Botterill and colleagues (2017). Their platform models each vine in 3D, lets an AI system decide which canes to prune, and cuts with a six-axis arm.
• Archie Jnr (Williams, MacDonald and colleagues, IROS 2024). It pruned 71.1% of 311 canes that required pruning, across 25 vines of a commercial vineyard, using three-cane pruning.
• Bumblebee (Carnegie Mellon and Cornell; Silwal et al., 2022). The one research trial of this job in the literature reviewed for this brief. Working behind a VMech pre-pruner on 20 vigorous Concord vines, it pruned 87% of prunable canes at 213 seconds per vine from two sides.
• Gebrayel and colleagues (2026, DOI: 10.1109/LRA.2026.3664530). A servoing-based system that pruned seven vine stocks, including outdoors in strong wind.

Commercially, Robotic Perception lists an autonomous pruner for pre-order, and Vision Robotics describes its grapevine pruner as a prototype dependent on financing.

Learning is competitive in the lab; no field head-to-head exists in the literature reviewed.

• In an indoor laboratory, a learned vision controller with force control raised pruning success from 46% to 77% over 26 trials each, against vision-only control (You et al., 2022).
• On a laboratory proxy tree, a learned visuomotor policy succeeded in 7 of 20 trials, against 2 of 10 for a point-cloud planner, a difference these sample sizes cannot distinguish (Fisher exact p = 0.68). In simulation, the same policy reached half the success of an oracle planner (Jain et al., 2025).

The gap is attended cost, including capital. Bumblebee's 213 seconds per vine, at the tested block's 537 vines per acre, implies 31.8 robot run-hours per acre. Its own cost model assumed 18 run-hours and counted neither an attendant nor capital. Other varieties and densities are unmeasured in the literature reviewed for this brief.

3. Foundational Research

Botterill T, Paulin S, Green R, Williams S, Lin J, Saxton V, Mills S, Chen X, Corbett-Davies S (2017). "A Robot System for Pruning Grape Vines." Journal of Field Robotics, 34(6), 1100-1122. DOI: 10.1002/rob.21680.

• Method: a row-straddling platform builds 3D vine models from trinocular stereo by incremental bundle adjustment. An AI system decides which canes to prune, and a six-degree-of-freedom arm plans cuts with a rapidly exploring random tree planner.
• Results: trajectory error below 1% over a 96 m row of 59 vines at 0.25 m/s. The arm cut 8.4 canes per vine on average and planned in 1.5 s per vine, and pruning took 2 minutes per vine in field trials. The abstract does not state how many vines were pruned.
• Why it matters: planning latency is not the bottleneck in an integrated vine pruner.

Williams H, Smith D, Shahabi J, Gee T, Qureshi A, McGuinness B, Harvey S, Downes C, Jangali R, Black K, Lim H, Duke M, MacDonald BA (2024). "Archie Jnr: A Robotic Platform for Autonomous Cane Pruning of Grapevines." 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 11736-11743. DOI: 10.1109/IROS58592.2024.10802290.

• Method: a vision system assesses vine structure, and the platform prunes the lower-quality canes as an expert pruner would. It was evaluated in a commercial vineyard under a three-cane method.
• Results: 71.1% of the 311 canes that required pruning were pruned successfully, across 25 vines.
• Why it matters: selective cutting in commercial vineyards is not a single-group result, and 25 vines is the largest sample of pruned vines among the cited trials that report one.

Silwal A, Yandun F, Nellithimaru A, Bates T, Kantor G (2022). "Bumblebee: A Path Towards Fully Autonomous Robotic Vine Pruning." Field Robotics, 2, 1661-1696. DOI: 10.55417/fr.2022051.

• Method: twenty vigorous Concord vines in one commercial-vineyard row were pre-pruned with an OXBO VMech pre-pruner. A ground robot with a controlled-lighting camera system and a redundant arm then spur-pruned them as a follow-up pruner, to a simplified rule of four buds per cane.
• Results: 87% of prunable canes pruned (83 of 95), at 213 s per vine from two sides and 137 s from one side. Bud detection was 95% and pruning-point selection 94% accurate.
• Why it matters: follow-up robotic pruning has worked once in the field, on one variety. The paper also supplies the per-acre cost baseline, and its measured cycle time disagrees with its modeled run time.

You A, Kolano H, Parayil N, Grimm C, Davidson J (2022). "Precision fruit tree pruning using a learned hybrid vision/interaction controller." 2022 IEEE International Conference on Robotics and Automation (ICRA), 2280-2286. DOI: 10.1109/ICRA46639.2022.9811628.

• Method: a Proximal Policy Optimization vision controller, trained only in simulation, aligns the cutter. An admittance controller on a 500 Hz force-torque sensor then completes the approach.
• Results: in 26 indoor laboratory trials each, the hybrid succeeded 77% of the time, against 46% for vision-only control and 19% and 23% for open-loop position control.
• Why it matters: contact-aware control lifts cut success substantially, which bears on follow-up cuts close to cordon wire.

You A, Parayil N, Krishna JG, Bhattarai U, Sapkota R, Ahmed D, Whiting M, Karkee M, Grimm C, Davidson J (2023). "Semiautonomous Precision Pruning of Upright Fruiting Offshoot Orchard Systems: An Integrated Approach." IEEE Robotics and Automation Magazine, 30(4), 10-19. DOI: 10.1109/MRA.2023.3309098.

• Method: perception, pruning-point selection and manipulation were integrated and field-tested in a planar sweet cherry orchard.
• Results: 58% cutting success across 10 trees. The authors' preprint reports:
  • 22 of 38 detected targets cut; the cut losses were 6 planning failures and 10 reach failures.
  • 115 false positives, which the authors call the biggest issue; cuts on the 96 that were spurs were withheld by hand.
  • 27 missed intersections, 25 of them from implementation issues in the intersection algorithm.
  • 284 s to scan a region.
• Why it matters: in an unattended system every false positive would cost an arm cycle or a wrong cut. Detection precision is therefore a supervision variable as much as an accuracy one.

Jain A, Grimm C, Lee S (2025). "Learning to Prune Branches in Modern Tree-Fruit Orchards." 2025 IEEE International Conference on Robotics and Automation (ICRA), 15553-15559. DOI: 10.1109/ICRA55743.2025.11128361.

• Method: a Proximal Policy Optimization policy maps wrist-camera optical flow to cutter motion at 2 Hz. It was trained in simulation and transferred zero-shot.
• Results: in 3,000 simulated trials, 30.06% success against 60.4% for an oracle planner with perfect geometry. On a laboratory proxy tree, 7 of 20 real trials (35%), against 2 of 10 for a point-cloud RRT-Connect planner.
• Why it matters: learned control without full reconstruction is competitive at small sample sizes, which a field trial has not yet tested.

Navone A, Martini M, Chiaberge M (2025). "Autonomous robotic pruning in orchards and vineyards: A review." Smart Agricultural Technology, 12, 101283. DOI: 10.1016/j.atech.2025.101283.

• Method: a review of robotic pruning literature from 2014 to 2024, organized by perception, skeletonization and control.
• Results: manual pruning runs up to 25% of annual labor cost in fruit production. The abstract does not state the number of works reviewed or the source of that figure.
• Why it matters: the field has matured component by component, while complete-system economics remain open.

4. Competitive Landscape

Market definition for the count: machines sold in the United States that select and execute individual spur- or cane-level dormant cuts on grapevines without a person choosing each cut.

Company	Offering	Funding or price	Status	Counted
Robotic Perception (Israel)	Autonomous pruner, vineyards and orchards	Not disclosed	Listed for pre-order	Yes, pre-order
PeK Agroline (Slovenia)	Slopehelper platform, robotic branch cutting with a scissor manipulator	Not disclosed	Marketed; selectivity and US sale unverified	Unresolved
Vision Robotics Corp	Grapevine pruner, bilateral cordon	Not disclosed	Prototype, dependent on financing	No
Pellenc	VISIO pre-pruner; Precision Pruner with cordon tracking	Not disclosed	Sold	No, cuts to a line
Westside Equipment Co.	VMECH Chariot, pre-pruning and near-finish pruning	Acquired VMECH June 2023	Sold	No, cuts to an envelope
OXBO	VMech pre-pruning head	About $30,000 (Silwal estimate)	Sold	No, cuts to an envelope

The count is zero on confirmed general sale, one on pre-order and one unresolved.

Funded research programs, and one commercial fast follower, matter more than the pruner companies.

• Carnegie Mellon. A $1.0 million USDA NIFA award, which ended in August 2026, proposed combining learning methods with classical approaches to prune vines in commercial fields.
• University of Auckland. New Zealand's MaaraTech programme was reported by Rural News Group as a NZ$16.8 million, five-year MBIE Endeavour Fund project from 2018, led by Auckland. It described a planned prototype carrying robotic arms for pruning and thinning; the five-year programme began in 2018.
• Orchard Robotics (commercial, not research). It raised a $22 million Series A in September 2025 for tractor-mounted crop imaging. It already holds a plant-level data layer and is a plausible fast follower.

Why the space is not commoditized. Three barriers compound:

• Throughput and supervision. Pre-pruners sell on acres per hour. A selective robot on the tested block needs about 31.8 run-hours per acre, several times the capital-inclusive breakeven set out in Section 6.
• Operation. Dormant vineyards are shared with tying and brush crews, so employee exposure is reasonably predictable and the practical machine is attended.
• The season clock. Each field iteration needs a dormant season, a clock that southern-hemisphere and multi-site trials can shorten but not remove.

What commoditization would require: attended cost per acre, including capital, below the $378 hand follow-up across several varieties, sold through an equipment channel; on a two-season validation clock from 2026, that looks unlikely before 2029 (an estimate, not a forecast).

5. Total Addressable Market

Market definition: annual United States labor spend on the selective hand follow-up cut after mechanical pre-pruning.

Input	Value	Source
Bearing grape acreage, 2025	883,000 acres	USDA NASS, May 2026
Hand follow-up after pre-pruning	18 labor hours per acre	Silwal et al., 2022
Loaded labor rate	$20.30 to $23.75 per hour	UC Davis, 2021; WSU TB106E, 2024

• Follow-up cost per acre: 18 h x $20.30 to $23.75 = $365.40 to $427.50
• TAM, upper envelope (all bearing acreage pre-pruned and hand-finished): 883,000 x $365.40 to $427.50 = $323 million to $377 million per year

The envelope errs in both directions. It overstates fully mechanized acreage, and it misclassifies acreage still hand pruned without a pre-prune, as in the UC Lodi budget. The 2021 wage also understates current cost.

Top-down cross-check. MarketsandMarkets estimates the global agricultural robots market at $17.73 billion in 2025, reaching $56.26 billion by 2030 at a 26.0% CAGR (November 2025). Its segmentation leads with unmanned aerial vehicles and milking robots. The US follow-up envelope is about 2% of the global figure.

SAM. The serviceable market starts where the field evidence applies. It is defined by variety class, and it assumes every acre in it buys a hand follow-up, which growers who prune fully mechanically do not.

Segment	Bearing acres	Follow-up envelope per year	Status
Initial: Washington juice plus all New York acreage	17,000 + 33,000 = 50,000	$18 million to $21 million	Near-term
Expansion: vinifera wine grapes (California, Washington, Oregon)	510,000 + 50,000 + 32,000 = 592,000	$216 million to $253 million	After follow-up spur pruning is proven on vinifera cordons

NASS publishes no type split for New York. The initial SAM is under $100 million and is flagged as such.

Payment and reimbursement pathway. No CPT or HCPCS code applies; this is agricultural equipment. Growers pay by capital purchase or by per-acre service.

• Capital purchase. Internal Revenue Code section 179 allows expensing up to $2,500,000 for tax years beginning in 2025 and $2,560,000 in 2026. Property acquired after January 19, 2025 qualifies for 100% bonus depreciation (IRS Instructions for Form 4562; P.L. 119-21).
• Per-acre service. The price ceiling is the $378 per acre hand follow-up.

6. Research Gap and Commercial Opportunity

Follow-up robotic pruning has worked once in the field. The opportunity is a follow-up pruner whose attended cost per acre, including capital, beats the crew across varieties, and that has not been built.

The cost model the published robot omits. Silwal's robot follow-up costs $80.64 per acre, which is fuel and lubrication over 18 run-hours ($4.48 per hour), or $161.90 combined with the pre-prune. It counts no attendant and no capital. Adding both gives:

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

The attendant rate is the paper's own $21 per hour. The capital term uses assumptions stated here, not measured:

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

Robots per attendant	Before capital	Including capital	Gain needed from 31.8 run-hours
1	14.8	8.5	3.7 times
2	25.2	11.2	2.8 times
3	32.9	12.5	2.5 times
3, at half the unit cost	32.9	18.1	1.8 times

Supervision ratio is rule-bound. Under California's Title 8 section 3441(b) as written, each self-propelled platform carries its own attendant where employees are exposed, so the one-robot row applies there: 8.5 run-hours per acre, a 3.7-fold gain. Two or three units per attendant apply only where rules permit (Washington and New York rules were not researched for this brief), under a variance, or when several arm modules share one operator-carrying vehicle, with capital then counted per vehicle. A shorter window raises the bar: at 480 run-hours per season, as in a February-only window, three units per attendant would need 6.4 run-hours per acre.

Three coupled targets decide the product:

• Supervision ratio. How many units one attendant can oversee, which depends on intervention rate.
• Run-hours per acre.
• Manufactured unit cost.

In the literature reviewed, no published system has measured the first, and none has attempted the third.

The integration gaps behind those targets:

1. Intervention rate. Supervision ratio depends on how often a unit needs a person. In the orchard field trial, the authors withheld cuts on 96 of 115 false detections by hand, which is exactly the intervention load an attendant would carry.
2. Multi-arm cut scheduling. Published systems use one arm per side. Two arms pruning both sides at once cannot finish a vine faster than the 137 seconds one side took, still about 20.4 run-hours per acre (a derived bound). The proposed approach puts more arms per side and assigns cuts across them per vine segment as a mixed-integer linear program minimizing platform dwell time.
3. Variety coverage. In the literature reviewed for this brief, follow-up spur pruning after a pre-prune has been measured only on Concord. Selective cane pruning has been field-tested separately.
4. Correctness and outcome. The follow-up trial used a simplified four-bud rule rather than professional pruner labels, and no reviewed study reports next-season yield after robotic pruning.
5. Manipulation choice under field conditions. Classical planning with position-based visual servoing is the baseline, and a Proximal Policy Optimization contact-aware policy is the challenger.

Why incumbents have not closed it. Pellenc, Westside and OXBO sell acres per hour, which a selective cutter at research cycle times does not fit. The lead US research lineage ran on episodic grants from 2015 to 2026, and its prototype carries a $115,000 capital estimate with no cost-down path. Venture capital went to imaging, which needs no attended safety case.

This is integration, supervision and manufacturing engineering, not unsolved science.

7. Comparable Funded Projects

The seven awards below commit $19.8 million since 2020 to robotics and automation of specialty-crop production operations. Two of them are specific to pruning.

Funder and program	Recipient and PI	Award	Amount	Period	Relevance
USDA NIFA AFRI	Carnegie Mellon University	2021-67021-35974	$1,000,000	2021 to 2026	Autonomous grapevine pruning; on topic
USDA NIFA AFRI	Oregon State University	2020-67021-31958	$374,704	2020 to 2024	Autonomous dormant pruning of fruit trees; on topic
USDA SCRI	Cornell University	2026-51181-46280	$5,835,164	2026 to 2030	Multi-purpose tree fruit robot, harvesting and thinning
USDA SCRI	North Carolina State University	2024-51181-43291	$9,825,677	2024 to 2029	Nursery labor automation including pruning
NSF Cyber-Physical Systems	Washington State University, Ming Luo	2312125	$1,199,998	2023 to 2027	Human-robot apple harvesting
NSF Cyber-Physical Systems	University of Kentucky, Biyun Xie	2437812	$1,179,584	2025 to 2028	Autonomous robotic tomato phenotyping in large-scale greenhouses
NSF with USDA NIFA	Michigan State University, Zhaojian Li	2540586	$400,000	2026 to 2029	Occlusion-resilient aerial robotic apple harvesting

Principal investigators for the USDA awards are not published in federal spending records. Cornell's award record cites the Specialty Crop Research Initiative priority on mechanization and automation to reduce manual labor, and names a robotics company partner for commercialization. Both pruning-specific awards have now ended.

8. Opportunity Assessment

TRL by component

Component	TRL	Evidence
Selective follow-up spur pruning on pre-pruned vines, one arm per side	5	20 Concord vines, one commercial-vineyard row (Silwal et al.)
Selective cane pruning, full vine	5	25 vines, 311 canes, commercial vineyard (Williams et al.)
Learned pruning manipulation	4	Laboratory only
Attended multi-arm operation at a capital-inclusive breakeven	3	Unmeasured
Follow-up spur pruning on vinifera cordons against professional labels	3	Unmeasured

The composite TRL for the product configuration is 4. Reaching TRL 5 at product scope requires field measurement of attended multi-arm follow-up pruning, with intervention rate, on at least two varieties.

Top technical risks

1. Intervention rate caps supervision ratio. Mitigation: log every attendant intervention by cause from the first dormant season, and gate scale-up on intervention time below one third of run time.
2. Cycle latency on a moving platform. Reconstruction, decision and cut must fit within dwell time per vine segment. Mitigation: stop-and-go dwell with scan-once reconstruction, plus a measured per-cut latency budget.
3. Unit cost does not fall. Mitigation: a bill-of-materials gate tied to the capital-inclusive breakeven from the first prototype.

Regulatory pathway

• FDA: no clearance applies.
• California Title 8, section 3441(b): self-propelled equipment in motion must have an operator stationed at the vehicular controls. The operator may be at a location on the vehicle other than the driving position, provided starting, accelerating, decelerating and stopping controls are there. Steering controls are also required where the machine steers other than by ground or furrow, or exceeds 2 mph. The attended design therefore puts an attendant on each platform, which under the rule as written means one attendant per platform.
• Cal/OSHA memorandum, August 30, 2024: it states that autonomous agricultural vehicle use at a worksite with no employees and no employee access is not a Title 8 violation. It rescinds Cal/OSHA's opposition to autonomous agricultural vehicles and proposes an advisory committee on vehicles under 500 lb, under 20 horsepower and under 2 mph. That class bears on a future unattended configuration.
• Petition 596: the Standards Board denied this petition for driver-optional tractors in 2022.
• Regulatory precedent: Cal/OSHA granted Monarch Tractor an experimental variance on August 6, 2021, to run autonomous tractors in two fields for five years, initially with an operator at the controls.
• Other states: Washington and New York operator requirements were not researched for this brief, and attended operation there is a safety design choice.
• Product precedent: John Deere revealed a production autonomous 8R tractor on January 4, 2022.
• Machinery safety and Europe: ISO 18497-1 through 18497-4 (2024) apply. Regulation (EU) 2023/1230 applies from January 20, 2027, with third-party conformity assessment where learning AI modules ensure safety functions.

Locked versus adaptive algorithms. The product ships frozen perception and cut-selection models per dormant season, retrained off-machine and verified against a fixed benchmark, with deterministic safety functions that keep a European release outside the self-evolving safety-function trigger. FDA's Predetermined Change Control Plan guidance has no jurisdiction here and is cited only as a model of a pre-specified update envelope.

Regulation as a time moat, with limits. An attended safety case, ISO 18497-4 verification history and proprietary multi-season field data accumulate with each dormant season. Southern-hemisphere and multi-site trials can narrow that clock, and the open benchmark is shared by design, so the advantage is a head start measured in seasons rather than a barrier capital cannot cross.

9. Team Requirements

Executing this opportunity requires four capability areas:

• Field robotics and learned control. Multi-arm manipulation, cut scheduling, contact-aware control, and fair benchmarks of learned against classical methods.
• Sensing and experimental design. 3D vine reconstruction, high-precision spur and cane detection, intervention logging, and randomized field trials linked to next-season outcomes.
• Manufacturing engineering. Design for manufacturability of arms, cutters and a row-straddling platform at a unit cost the capital-inclusive breakeven allows.
• Viticulture and grower access. Pruning rules across varieties and trellis systems, plus commercial vineyard sites for dormant-season trials.