Cyber-Physical System Design Space Exploration for Affordable Precision Agriculture

This is an authors’ copy of the paper to appear in the 2026 Design, Automation & T est in Europe Conference (D A TE). Cyber -Physical System Design Space Exploration for Af fordable Precision Agriculture Pa wan Kumar Arizona State University pkumar97@asu.edu Hokeun Kim Arizona State University hokeun@asu.edu Abstract —Precision agriculture promises higher yields and sustainability , b ut adoption is slowed by the high cost of cyber - physical systems (CPS) and the lack of systematic design meth- ods. W e present a cost-aware design space exploration (DSE) framework for multimodal drone–r over platf orms to integrate budget, energy , sensing, payload, computation, and communica- tion constraints. Using integer linear programming (ILP) with SA T -based verification, our approach trades off among cost, coverage, and payload while ensuring constraint compliance and a multitude of alternatives. W e conduct case studies on smaller and lar ger-sized farms to sho w that our method consistently achieves full coverage within budget while maximizing payload efficiency , outperf orming state-of-the-art CPS DSE approaches. Index T erms —Design space exploration, Precision agricultur e, Cyber -physical systems, Cost-aware optimization I . I N T RO D U C T I O N Modern agriculture faces challenges to sustainably feed a growing global population amid workforce shortages, climate variability , and constrained natural resources [1]. Precision agriculture has emerged to tackle these issues with adv anced sensing, automation, and data-driven management to optimize crop production. In particular , cyber-physical systems (CPS) integrate with sensors, microcontrollers, and real-time image processing, enabling continuous, non-destructiv e monitoring of crop growth and targeted interventions to boost yields while keeping hardware costs moderate [2], [3]. Such CPS-driv en precision agriculture not only improves resource use efficiency (e.g., minimizing usage of water , fertilizer , and pesticide) but also helps offset labor constraints and environmental impacts, allowing farmers to achie ve higher yields more sustainably [4]. Design space exploration (DSE) is critical for dev eloping cost-effecti ve CPS in precision agriculture, b ut it is challenging due to conflicting goals of minimizing cost while ensuring adequate coverage, payload, and runtime [4], [5]. Balancing these trade-offs requires systematic methods to explore hard- ware–software configurations under budget and operational constraints. Recent work emphasizes multimodal platforms that maximize field co verage and enable tar geted interventions for sustainability and reduced chemical usage [6], [7]. This synergy between DSE and multimodal platforms provides a scalable pathway to practical, data-driv en precision agri- culture. Figure 1 shows an integrated multimodal platform hardware configuration, illustrating the trade-offs among size, cov erage, and computing capacity to be ev aluated. A. Related W ork and State-of-the-Art Appr oaches Recent studies have introduced cost-effecti ve CPS proto- types and predictive modeling to non-destructiv ely estimate Sm a ll UAV M e d iu m UAV Loc al edge se r ve r C omputat i o n offl oadi n g O n - dev i c e c omputi ng Large U G V sm a ll UG V O n - dev i c e c omputi ng Fig. 1: Example illustration of DSE for cost-effecti ve multimodal software- hardware platforms with unmanned ground vehicles (UGVs or rovers) and unmanned aerial vehicles (U A Vs or drones) for precision agriculture. plant growth, thereby promoting sustainability and improving crop yields [2], [8]–[10]. Doering et al. [11] propose a U A V - based platform with adapti ve mission planning for precision agriculture. These methods integrate CPS in precision agricul- ture to facilitate data acquisition and adaptive decision-making in response to dynamic en vironmental conditions. DSE is crucial in optimizing CPS configurations by balanc- ing performance, cost, and computational efficienc y . M ¨ uhleis et al. [12] integrate control performance directly into DSE. Fitzgerald et al. [13] develop a toolchain unifying discrete and continuous CPS models. Herget et al. [14] improve workload modeling for distributed CPS. Liao et al. [15] refine subspace pruning to cut computational overhead. Xiao et al. [16] pro- pose ContrArc, which uses contract-based pruning to eliminate infeasible architectures before exploration. Y u et al. [17] explore symbiotic CPS designs using human-in-the-loop feed- back to iterativ ely refine UA V architectures. Zheng et al. [18] dev elop a gradient-free framework for cross-domain U A V de- sign exploration. There is, ho wev er , a critical gap that none of these methods target CPS DSE for precision agriculture. This motiv ates the need for a domain-aware CPS DSE frame work tailored to the constraints and goals of agricultural automation. B. Resear ch Pr oblem and Contributions A unique challenge in the agricultural CPS DSE problem is the co-parameterization of crops and applications, with non-linear relationships between crops and sensing/coverage requirements. Specifically , dev eloping CPS for agriculture re- quires balancing competing objectives: minimizing cost while ensuring sufficient sensing, computation, and field coverage. Current state-of-the-art methods are limited to optimizing individual factors, lea ving farmers without a systematic way to ev aluate full platform designs within budget and constraints. W e address this gap with a cost-aware DSE framew ork that ev aluates heterogeneous multimodal platform configurations across budget, energy , and sensing constraints. The underlying integer linear programming (ILP) formulation is non-trivial because many of these relationships are inherently non-linear , requiring careful modeling and relaxation to obtain feasible solutions. T o guarantee feasible outputs, we complement our approach with SA T -based verification to check the generated platforms and filter out inv alid ones. Our framew ork produces alternativ e designs that achiev e high coverage and sensing performance at the lowest possible cost, providing a principled method to navigate complex design trade-offs. W e release the design, implementation, data sets, and ev aluation results of our approach through an open-source repository . 1 I I . P RO P O S E D A P P R OA C H This section outlines our proposed DSE method with ex- planations of user inputs and hardware factors to be explored, leading to feasible platforms for agriculture. A. User Inputs and Design Considerations Users first provide key inputs that shape the design space: the total av ailable budget, the size of the farm area, the target crop type, and the desired applications, as shown below . 1. Budget: Sets an upper limit on the total allow able mon- etary cost for hardware configurations, ensuring that our platforms are practical for farms with varying financial constraints. 2. Farm area size: Determines the minimum area coverage requirement that the platform must achiev e. Larger farms may demand multiple vehicles or higher-capacity compo- nents to maintain efficienc y . 3. T arget crop: Determines the suitability of drones, ro vers, or both for field operations. Indoor crops grown in controlled en vironments typically fa vor rov ers, as drones are imprac- tical in confined spaces. Densely planted crops like fiber and legumes increase the risk of ground vehicle damage, making drones a better choice. Similarly , early-stage paddy fields are often waterlogged, limiting rover use and fav oring aerial monitoring. In contrast, crops such as cereals, trees, orchards, vines, vegetables, forage, and oilseeds benefit from hybrid drone–rov er platforms, combining aerial sur- ve ys with ground-lev el intervention. 4. Application: Defines exact tasks that the platform must perform. T ABLE I lists v arious agricultural applications that users can choose from when setting up our DSE frame work. Each application has a recommended processing mode, either off-board (using a local edge server) or onboard (computing directly on the vehicle), and specifies whether it is better suited to drones, rov ers, or both. Applications requiring control and actuation are handled onboard, while others are off-board. Once these inputs are defined, our DSE framework system- atically translates them into the design considerations that determine which combinations of chassis, motors, batteries, computing units, and sensors can meet the farm’ s needs within the budget and platform constraints. It then ev aluates each candidate using fiv e key metrics: 1 https://github .com/asu-kim/cps-dse-apa Application Computation Mode Platform General crop monitoring Off-board Both Thermal imaging Off-board Both Image stitching Off-board Both Soil monitoring Off-board Rov er Y ield estimation Off-board Both Quality control Off-board Rov er Autonomous fruit/vegetable picking Onboard Rover Mechanical weeding Onboard Rover Soil pH sampling Onboard Rover Climate mapping Off-board Rov er Fence/infrastructure inspection Off-board Drone Liv estock monitoring Onboard Both Beehiv e inspection Onboard Rover Frost & pest early-warning systems Onboard Both Fertilizing Onboard Rover T ABLE I: Overvie w of precision agriculture applications and their recommended processing modes and hardware platform compatibility with rover , drone, or both. 1. Cost: Sum of base vehicle, component (e.g., LiD AR, ma- nipulators), and edge-server costs. 2. Coverage: Maximum operational cov erage area (m 2 ), cal- culated by estimating av ailable battery energy after ac- counting for peak draws from motors, sensors, and onboard processors, then integrating over average platform speed. 3. Payload capacity: Mass (kg) available for additional com- ponents such as cameras, sensors, or manipulators, varying according to the target crop and application. 4. Runtime: T ime spent (hours) before recharge, computed under the same maximum-component-load assumption. 5. Crop–A pplication Parameterization: Parameters tailored to specific crop types and application scenarios. Our DSE frame work encodes b udget, area co verage, runtime thresholds, payload, communication-range requirements, and crop–application specifications into ILP . The goal of the ob- jectiv e function in Eq.1 is to minimize a weighted combination of normalized cost, area coverage, and payload. By varying priority weights ( α, β , γ ), our ILP solver systematically gen- erates all the possible drone-rov er configurations that pass all constraints. Then, it screens them for feasibility using the crop–application platform compatibility and component power limits, min x N X i =1  α ˜ C i x i + β ˜ A i x i − γ ˜ P i x i  (1) N denotes the total number of possible configurations, and i is the index of each configuration. For configuration i , ˜ C i is the normalized cost, ˜ A i is the area coverage, and ˜ P i is the per-platform payload. The variable x i indicates the number of units selected for configuration i . Finally , α , β , and γ are weight factors used to balance cost, area co verage, and payload capacity using the ROC weighting scheme described in III-A. B. Design Space Exploration In this section, we formalize the framew ork requirements through a series of linear inequalities that encode b udget limits, coverage and payload needs, runtime targets, com- ponent sizing rules, and communication-range expectations. These constraints approximate relationships that behave non- linearly in practice, such as the way weight influences energy consumption, how motor torque scales with usable payload, and how coverage depends on battery discharge and geometric range. T ogether, they define the feasible design space that our ILP solver searches to find valid drone-rover platforms. W e use equation (Eq. 2) to ensure the total cost does not exceed the user’ s budget, ( C i + extr a cost ) · x i + edg e cost ≤ B (2) Here, C i is the base monetary cost of configuration i , which includes the vehicle chassis, core hardware, and essential built-in components. B denotes the total user-defined budget. The term extr a cost accounts for additional monetary cost for application-specific parts, while edg e cost represents the monetary cost of the local edge server . Eq. 3 ensures selected rov ers and drones cov er at least the required farm size, A i · x i ≥ S (3) Here, S denotes the farm size specified by the user , x i is the number of units selected for configuration i , and A i represents the cov erage area each unit of configuration i can achiev e. T o determine how much area each selected configuration can cover , we use Eq.4 to find the operational coverage area for any drone or rover configuration in our DSE, A i = B AT T i · T i · W component ,i · f weight ,i ( W max ,i ) (4) Here, B AT T i and T i denote the number of batteries and tires in configuration i . W component ,i is the weight calculated in Eq. 8 W max ,i is the maximum payload of configuration i calculated in Eq. 7, and f weight ,i ( W max ,i ) captures the efficienc y reduction due to weight, as defined in Eq. 5. f weight ,i ( W max ,i ) = 1 1 + W max ,i / 100 (5) W e use Eq. 6 to find the total force the motors generate, which determines how much weight the platforms can mov e, W carry = τ × nm T R · g (6) Here, W carry is the effecti ve motor weight (in kg), computed by dividing the total torque-based force τ by the tire radius T R , and the gravitational acceleration g . The term nm denotes the number of motors in configuration i . For drone configurations, tire radius is omitted from the computation of W motors . W e use Eq. 7 and Eq. 8 to find the maximum payload of configuration i . W max,i = W carry ,i − W component ,i (7) W component ,i = W chassis ,i + W motor ,i + W battery ,i + W tires ,i (8) Here, W component ,i is the sum of physical weights. W chassis ,i , W battery ,i , W motors ,i , and W tires ,i represent the physical weights of the chassis, battery , motors, and tires, respecti vely of configuration i . For drone configurations, tire weight is omitted from the computation of W component . W e use the follo wing equation, Eq. 9, to ensure each selected drone or rover configuration includes at least one battery module to meet its energy requirements. X s B AT T i,s ≥ 1 , ∀ i (9) Here, B AT T i,s is a binary variable indicating whether battery option s is included in configuration i , and s represents the av ailable battery choices for rover configurations. Eq. 10 ensures each motor matches the chassis size, main- taining power -to-weight balance for reliable operation. M size ,i = C H size ,i , ∀ i (10) Here, M size ,i and C H size ,i denote the motor size and chassis size, respectiv ely , for configuration i . W e observe that drone configurations typically have a much shorter flight time than rovers due to higher power draw and limited battery capacity . T o balance this gap and ensure that drones can perform meaningful tasks without frequent recharging, we add Eq. 11 below as a flight time condition. f d ≥ F min (11) Here, f d denotes the estimated runtime (flight time in hours) of a specific drone configuration d , and F min is the minimum required flight time per unit, set to 0.2 hours. This threshold reflects the flight time of commercial drones such as the Flyability Elios 3 [19] and XAG V40 [20], which typically operate for about 0.2 hours. W e ensure the drone-rover configurations maintain full communication cov erage across the farm. Eq. 12 below cal- culates the number of cells 2 by dividing the farm area by the communication method’ s effectiv e range. l = √ S R comm × 1000 (12) Here, S is the total farm size specified by the user in m 2 (square meter), R comm is the range of a communication method (e.g., LoRa, cellular, W i-Fi) in km, and l is the number of cells (minimum of 1 cell) required to cov er the entire area. Eq. 13 then multiplies the number of cells from Eq. 12 by the cost per cell to compute the total communication cost. C comm = l × C cell (13) Here, C cell denotes the cost of a single communication cell, and C comm represents the total communication cost. A suitable computing unit is selected based on the ap- plication’ s processing requirements as shown in (T ABLE I). On-board processing requires real-time tasks such as image processing or autonomous navigation, so we assign a GPU or TPU. For off-board processing, the configuration may instead use a CPU, GPU, or TPU. Computation performance is ev aluated using Geekbench 5.4.1 [21]–[23] multi-core scores, which are deriv ed from workload execution time, the lower the ex ecution time higher the score. C. SA T -based V erification T o ensure the correctness of the ILP ev en after modification of the ILP formula by the users, we verify that each configu- ration meets the total cost and total coverage constraints using a SA T solver , PySA T [24]. W e encode feasibility as Boolean 2 W e use the term “cell” to refer to the area covered by each device of the communication method, analogous to cells in cellular networks. variables, enabling efficient SA T solving. The solver checks conditions in Eq. 14. total cost ≤ B , total cov erage ≥ S (14) where B is the budget and S the required farm size. Hardware components (motors, batteries, sensors, frames) contribute to these totals, keeping verification aligned with design. Similar SA T -based methods are used in many-core DSE [25], [26], validating our approach for CPS. I I I . E V A L UA T I O N T o assess the effecti veness of our approach, we first e valuate the design aspects considered in our approach. Second, we compare our approach’ s optimization method against state-of- the-art approaches. • Evaluation of design aspects (Section III-A) – W e con- duct an ablation study of the ILP-based multi-objectiv e optimizers: Pr oposed Appr oach ( PA ) (cost+ar ea+payload) , Ar ea+P ayload AP , Cost+Area CA , and P ayload+Cost ( PC ). • Optimizer model evaluation (Section III-B) – Com- parison of the proposed ILP approach against optimiza- tion methods existing state-of-the-art CPS DSE approaches including: simulated annealing , Bayesian optimization (BO) [17], random sear ch , genetic algorithm , portfolio optimization , discr ete searc h , Lengler [18], and PG-DSE with PSP and PEGA [15] as summarized in T ABLE III. Our ev aluation uses two farm settings as case studies: Case Study 1 : W e assume a smaller-scale farm with a farm size of 4047 m 2 (approx. 1 acre), which aligns with farm size distribution data published by the USD A NASS in the 2022 Census of Agriculture [27] and by the EU in 2020 [28], with a budget of $100K (USD), which aligns with entry-level costs for robotic platforms used in precision farming [29], [30]. W e select the tree crop (Section II-A) for its high monitoring demands and economic value, as sho wn in reports from the US and Europe, with the v alue of $2.9 billion USD in 2024 [31] for apple, and 1.17 million acres of farm land used for gro wing apples [32]. W e use autonomous fruit and vegetable picking (T ABLE I) as applications lev eraging CPU, GPUs, or TPUs for efficient harvesting tasks on rovers. Case Study 2 : W e assume a large-scale farm with a farm size of 40,469 m 2 (approx. 10 acres) [27], [28], with a budget of $1M (USD), which aligns with the cost for robotic weeding machines [29], [30], [33]. W e select vine (Section II-A) due to its high value and requirement for delicate handling, as shown in reports from the US and Europe, with the value of $6.2 billion USD in 2024 [31] for grapes, and 7.9 million acres of farm land used for growing grapes [32]. W e use general crop monitoring and yield estimation (T ABLE I), which uses off-board processing with CPU, GPU, or TPU. Across both ev aluations, we visualize scatter plots and compute a unified weighted score for each optimizer . A. Evaluation of Design Aspects Figure 2 illustrates trade-offs across cost, payload, and area cov erage for different optimization methods for Case Study 2. 1 0 0 0 50 3 . 5 K 4 . 0 K 4 . 5 K 5 . 0 K Un it Cos t in T ho us an ds ( $) (a ) 0 5000 10000 0 . 0 M 0 . 2 M 0 . 4 M 0 . 6 M 0 . 8 M 1 . 0 M 1 . 2 M 1 . 4 M T ot a l Cos t in M illio n s ( $) ( b ) 0 2 0 0 A re a c o v e r a g e ( m ²) 0 . 0 M 0 . 2 M 0 . 4 M 0 . 6 M 0 . 8 M 1 . 0 M 1 . 2 M T ot a l Cos t in M illio n s ( $) 4 06 0 0 4 08 00 4100 0 (c ) c o s t + are a p a y l oa d + c ost are a+ p ay l o ad P r op ose d A p p ro a c h Bu d g et C o ns t r ai nt F arm S i z e C ons t r ai nt Un it Pa y l o a d ( k g ) T ot a l Pa y l o a d ( k g ) Pro t o t yp e R o ve r Pro t o t yp e D r o ne 3 . 0 K Fig. 2: Trade-of fs among design aspects in Case Study 2 ($1M, 10 acres): (a) unit cost vs. unit payload (b) total cost vs. total payload, (c) total cost vs. area coverage. A unit refers to a single drone or rover instance, and a configu- ration may include multiple units to satisfy minimum f arm area cov erage. Fig. 2–(a) plots the unit cost vs. unit payload where, PA and CA yield balanced designs, whereas PC increases cost with increase in payload. AP produces a more scattered distribution. Fig. 2–(b) plots the total cost vs. total payload where, PC clusters near the origin, while PA and CA achie ve high payloads within $1M. Whereas, AP has lower payload and also exceeds the $1M. Fig. 2–(c) plots the total cost vs. area coverage where, PA and CA maximize area coverage without exceeding $1M (yellow shaded region). In contrast, AP attains minimum coverage but includes configurations that exceed $1M (red shaded region), whereas PC stays under $1M but fails cov erage (green shaded region). W eighted Score Design : W e prioritize cost to achiev e our primary goal, a cost-effecti ve DSE. W e assign area coverage the next highest priority so our framew ork can monitor the minimum required area for the applications. W e consider payload as the least critical factor once the drone–rov er carries the necessary sensors, as extra capacity adds little value. T o reflect this, we now use the Rank Order Cen- troid (R OC) [34] method to determine the final weights, following the approach demonstrated in the works [35]– [37]. Using ROC, we rank the criteria in order of im- portance and derive their normalized weights as follo ws: Design Aspects Case Study 1 $100K, 1 Acre Case Study 2 $1M, 10 Acres AP 0.404 0.377 PC 0.611 0.611 CA 0.417 0.389 PA 0.417 0.389 T ABLE II: W eighted-score comparison for design aspects in our case. • W eight for cost: 0.611 • W eight for area cov erage: 0.278 • W eight for pay- load: 0.111 Under these weights for Case Study 2, T ABLE II shows that the PC achiev es the highest score of 0.611, followed by PA and CA at 0.389. AP ranks lowest at 0.377. A similar pattern appears in Case Study 1. PA and CA again achiev e full area coverage within $100K, while PC stays with the highest weighted score but lacks full area coverage, and AP exceeds $100K despite minimum area cov erage. T ogether , we confirm that the PA and CA deliv er the best balance of cost and area cov erage across different farm sizes. T o summarize, Fig. 2 and T ABLE II confirm that the PA and Appr oach Domain Platform Evaluation Methodology DSE Formulation Optimizers Cost Breakdown DESTION 23 [17] General CPS UA Vs & Robot Car Simulation-based performance ev aluations include time & energy for robot car & four flight tasks for UA Vs Reward feedback & switching strategy Simulated Annealing, Bayesian Optimization (BO) Not av ailable DESTION 22 [18] U A V archi- tecture U A Vs only 4 mission-based benchmarks (hover , straight, circle, ov al) Graph-based design optimized by component selection or ac- tion sequences Discrete, Lengler , Portfolio, FastGA, Random Search Not av ailable ASP-D AC 23 [15] AD AS FPGA- Xilinx Artix-7 AD AS system: 10 PSMs, 52 MCCs, 244 MCC alternatives, & Synthetic systems with design space sizes PSMs le verage component-le vel MCCs to model time-triggered behavior PG-DSE with PSP and PEGA Not av ailable Proposed approach Precision agriculture Rovers & U A Vs Evaluation of design aspects & Optimizer model evaluation Minimize cost, maximize area coverage & payload Integer Linear Pro- gramming A v ailable T ABLE III: Comparison of the proposed and state-of-the-art CPS DSE approaches. 0 3 0 0 6 0 0 9 0 0 Pa y l o a d (K g ) 2 0 . 0 K 120. 0 K 100. 0 K 8 0 . 0 K 6 0 . 0 K 4 0 . 0 K T o t a l Co s t in T h o u s a n d s ( $) ( a) 0 9000 3000 6000 (b ) P r op ose d A p p ro a c h Simu la t e d Anne al i ng ( DES T I O N ' 2 3 ) B ay e s i an O pt i m i z a t i o n ( DES T ION ' 2 3 ) Ra n do m S ea rc h ( DES T I O N ' 2 2 ) Gen et i c A l go r i t h m ( DES T ION ' 2 2 ) PG - DS E ( AS P - DAC ' 2 3 ) D i s c ret e S ea rc h ( DES T ION ' 2 2 ) L en g l er ( DES T I O N ' 2 2 ) P o r tfo l i o ( DES T I O N ' 2 2 ) Bu d g et F arm S i z e A re a c o v e r a g e ( m ²) ( i )I ns uf f i c i e nt A re a C o v e ra g e ( ii) S u ffi c i e n t A re a C o v e ra g e Clu s t er Clu s t er Fig. 3: Comparativ e ev aluation of optimizers in Case Study 1 ($100K, 1 acre): (a) total cost vs. payload, (b) total cost vs. area coverage. CA achiev e full area coverage while keeping costs within the $1M budget in Case Study 2, which explains why they earn balanced weighted scores. Meanwhile, the PC scores higher due to its low cost and high payload, but it fails to fully meet the area coverage requirement. AP exceeds the budget ev en though it meets the area coverage target and ranks lowest in the weighted score. B. Optimizer Model Evaluation Fig. 3–(a) plots the total cost vs. payload for Case Study 1, where the $100K budget line appears as a blue dotted line. The PA spans a broad payload range while remaining below $100K, indicating flexible scaling across configurations. sim- ulated annealing concentrates around mid-to-high (approx. 400-800 kg) payload values within $100K. genetic algorithm and random search show a gradual cost increase as payload increases. discrete search sho ws only one configuration at 400 kg payload with a cost of 60K. BO shows an increase in costs without huge changes in payload. PG-DSE co vers a wide span, from light payload (approx. 50 kg) at mid cost (approx. $60K) to heavy payload (approx. 900 kg) at low cost (approx. $30K) with some exceeding $100K. Lengler and portfolio position themselves mainly around low payload (approx. 150 kg) and cluster together , shown in a dotted circle. Fig. 3–(b) plots the total cost vs. area coverage, where the farm size appears as a red dotted line, and the budget line appears as a blue dotted line. The PA sits close to the farm size line, sufficient area cov erage (Fig. 3–(b)-(ii)) with varied State-of-the-Art Optimizers and Pro- posed Optimizer Case Study 1: $100K, 1 Acre Case Study 2: $1M, 10 Acres Simulated Annealing (DESTION 23) 0.325 0.465 Bayesian Optimization (DESTION 23) 0.498 N/A Random Search (DESTION 22) 0.496 0.475 Genetic Algorithm (DESTION 22) 0.503 0.527 Discrete Search (DESTION 22) 0.307 0.600 Lengler (DESTION 22) 0.634 0.686 Portfolio (DESTION 22) 0.634 0.686 PG-DSE (ASP-DA C 23) 0.619 0.360 Proposed Approach 0.634 0.686 T ABLE IV: W eighted-score comparison for optimizer ev aluation in Case Study 1 and Case Study 2. costs. simulated annealing extends beyond the minimum farm size, reaching sufficient area coverage (Fig. 3–(b)-(ii)) with costs upto approx. 95K. BO and genetic algorithm show a similar trend where area cov erage increases with cost. random search achieves moderate area coverage of about 4100–8500 m 2 and stays under budget. discrete search shows only one configuration at approx. 10000 m 2 . PG-DSE e xplores a broader solution space using e volutionary operators, which can momentarily include configurations in insufficient area cov erage (Fig. 3–(b)-(i)). Lengler and portfolio cluster near the farm size line and have similar cost, showing a similar trend between these optimizers. T ABLE IV presents the weighted-score comparison for optimizer model ev aluation, using the ROC weighting scheme described in III-A. In Case Study 1, the PA , Lengler , and portfolio top the ranking by striking the best balance among cost, area cov erage, and payload. PG-DSE ranks next highest, driven by multiple low-cost configurations, though some solutions incur budget penalties or provide limited area cov erage. genetic algorithm claims the next highest scores, lev eraging strong area cov erage with controlled spending. BO , Random search and simulated annealing follow , as they identify feasible configurations but exhibit a higher cost for similar area cov erage or payload compared to genetic algorithm . Discrete search earns a slightly lower score due to limited designs. In Case Study 2, the PA , Lengler , and portfolio achie ve the highest overall score by effecti vely balancing payload, area coverage, and cost. discrete search follow at a distance, genetic algorithm with close behind, followed by random search and simulated annealing . PG-DSE scores the lo west due to budget penalties or limited area coverage. Also, BO from DESTION 23 [17] giv es no results since we lack hyper- parameters (e.g., optimization calls, acquisition settings) that were unavailable, forcing untuned, general-purpose runs that cannot uncov er viable configurations under tight constraints. The trends shown in Fig. 3 align well with the weighted scores in T ABLE IV. PA consistently shows strong trade-offs across cost, payload, and area coverage, explaining its top score in both weighted scores and plots. genetic algorithm , BO and random search follow a similar pattern by keeping costs balanced while reaching practical area co verage. discr ete search , and simulated annealing show moderate performance with lower scores. Although PG-DSE scores a high weighted score, it has configurations that exceed the budget or provide limited area coverage. Lengler and portf olio perform simi- larly , with overlapping plots and identical scores. C. V erification of DSE Results Method V erified by SA T T otal Configs V alid Configs In valid Configs P A 18 18 0 AP 20 17 3 CA 18 18 0 PC 13 0 13 Simulated Annealing 12 12 0 BO 20 20 0 PG-DSE 20 4 16 Random Search 20 20 0 Genetic Algorithm 20 20 0 Discrete 1 1 0 Lengler 18 18 0 Portfolio 4 4 0 T ABLE V: SA T -based constraint verification across optimization methods, for Case Study 1. T ABLE V summarizes the results of constraint verification by our SA T - based approach described in Section II-C for Case Study 1. The following approaches return only valid configurations without any in valid ones: PA , CA , BO , discrete , Lengler , random search , genetic algorithm , portfolio , and simulated annealing . In contrast, PC completely fails, returning all in valid results (13 out of 13), while AP and PG-DSE return 3 and 16 in v alid configurations, respectively , out of 20 configurations. This highlights the strength of our ILP-based optimizer and the importance of verification in eliminating infeasible designs. D. Execution T ime Analysis Our approach can run on commercial off-the-shelf personal computers, such as laptops. W e measure the ex ecution time of our approach on a computer with an AMD Ryzen 7 processor (8 cores @ 3.6 GHz), 16GB RAM, Windo ws 11, and Python 3.10. In Case Study 1, PA requires 2.5 s to execute. PG- DSE completes in approximately 487 ms, while both genetic algorithm and portfolio finish in about 10 ms. Each of Lengler , random search , simulated annealing , and discrete search takes up to 5 ms to complete. In contrast, BO takes the longest time at 35.3 s. Also, our SA T solver completes verification almost instantly , in just 8.6 ms. Platform Hardware Configurations Payload (kg) Runtime (hours) Cost (USD) Rover Plastic large body, large motor, medium battery, RPi 4B 34.67 123.23 3449.56 Drone metal body , large motor , large battery , RPi 4B 0.1 0.22 2345.34 T ABLE VI: Prototype hardware specifications deriv ed from Case Study 2 solver outputs, with the drone implemented using a carbon- fiber body instead of the solver -selected metal body . E. Har dware Pr ototype Implementation T o demonstrate the feasibility of our DSE approach for precision agriculture, we build and verify a prototype U A V - UGV platform derived directly from the solver outputs of the proposed method, as shown in Fig. 4 and T ABLE VI. P r ot ot y p e D r on e P r ot ot y p e r ov e r Loc al e d g e ser v er P las t ic lar g e b od y Lar g e t ir e s C ar b on f ib e r b od y Fig. 4: Our prototype hardware constructed in line with the proposed approach’s outputs. Our prototype con- sists of a ro ver , a drone, and a local edge server . The rover uses a Rasp- berry Pi (RPi) 4B and a large plastic chas- sis with tires, provid- ing sufficient runtime while supporting nec- essary sensors. The drone (Duckiedrone DD24-B [38]) follows the solver-selected configuration, except for the body . While the solver specifies a metal body , we use a carbon-fiber frame for the lightweight, while keeping all other hardware compo- nents unchanged. Also, a local edge server ($2,000) provides off-board computation. T ogether, these elements confirm that ILP+SA T outputs are not abstract platforms but realizable, low-cost agricultural systems. I V . C O N C L U S I O N In summary , we propose a cost-aware DSE framew ork for multimodal UA V–UGV platforms in precision agriculture. By formulating the problem as an ILP with SA T verification, we address the challenges of balancing cost, coverage, and payload under strict resource constraints, while outperforming existing methods in terms of feasibility , cost, and the number of alternativ e configurations. Limitations, scope, and future work of this paper are further discussed below . Limitations and Scope: T o optimize the ILP , our framework targets static, deterministic conditions. These choices stream- line the ILP but limit its ability to reflect soil conditions, weather patterns, and long-term degradation. Communication and sensing are modeled at a high lev el, which can miss interference effects or routing complexity in large farms. Linear approximations of non-linear behaviors such as bat- tery discharge, weight–efficiency decay , and torque–payload scaling further narrow realistic behavior . Future W ork: W e plan to extend our framework to other CPS domains with non-linear trade-offs, such as warehouse monitoring and disaster response. These settings introduce challenges like cov erage vs. communication, mobility vs. energy , and safety vs. latency . W e continue modeling these within an ILP and verifying feasibility with SA T . Although each domain requires its own constraint set and weighting, the same systematic exploration process remains, making this work a foundation for broader CPS optimization. A C K N O W L E D G M E N T This work was supported in part by the National Science Foundation (NSF) under grants #2449200 (NSF I/UCRC for IDEAS) and POSE-#2449200 (An Open-Source Ecosystem to Coordinate Integration of Cyber-Physical Systems). R E F E R E N C E S [1] J. McFadden, E. Njuki, and T . Griffin, “Precision agriculture in the digital era: Recent adoption on U.S. farms, ” United States Department of Agriculture, USD A Miscellaneous 333550, 2023, dOI: 10.22004/ag.econ.333550. [Online]. A vailable: https: //ideas.repec.org/p/ags/usdami/333550.html [2] P . Kumar and H. Kim, “Cost-effecti ve cyber-physical system prototype for precision agriculture with a focus on crop growth, ” in 2024 Interna- tional W orkshop on Rapid System Prototyping (RSP) . IEEE, October 2024, pp. 35–41. [3] P . Castillejo et al. , “ Aggreg ate farming in the cloud: the AFarCloud ECSEL project, ” Micr opr ocessors and micr osystems , vol. 78, p. 103218, August 2020. [4] A. Pretto et al. , “Building an aerial–ground robotics system for pre- cision farming: An adaptable solution, ” IEEE Robotics & Automation Magazine , vol. 28, no. 3, pp. 29–49, August 2021. [5] D. Madro ˜ nal, F . Palumbo, A. Capotondi, and A. Marongiu, “Unmanned vehicles in smart farming: A survey and a glance at future horizons, ” in Pr oceedings of the 2021 Dr one Systems Engineering and Rapid Simula- tion and P erformance Evaluation: Methods and T ools Proceedings , ser. DroneSE and RAPIDO ’21. Association for Computing Machinery , January 2021, pp. 1–8. [6] P . Kumar , Y . Dube, and H. Kim, “Cost-effecti ve rover for farms, ” in Pr oceedings of the ACM/IEEE 16th International Confer ence on Cyber- Physical Systems (with CPS-IoT W eek 2025) , ser. ICCPS ’25, May 2025, pp. 1–2. [7] United States Government Accountability Office, “Precision agriculture: Benefits and challenges for technology adoption and use, ” 2024, accessed: 2025-06-26. [Online]. A vailable: https://www .gao.go v/products/gao- 24- 105962 [8] S. Bouarourou, C. Kanzouai, A. Zannou, E. H. Nfaoui, and A. Boulaalam, “Crop yield prediction in IoT: A hybrid feature selection approach using machine learning models, ” in 2024 3r d International Confer ence on Embedded Systems and Artificial Intelligence (ESAI) . IEEE, December 2024, pp. 1–5. [9] H. Qiu, Z. Zhou, Y . Peng, J. Liang, J. Shang, X. Xi, and R. Zeng, “Multi-terminal cooperative AIoT smart agriculture system based on growth prediction model, ” in 2024 IEEE International Confer ence on Smart Internet of Things (SmartIoT) . IEEE, Nov ember 2024, pp. 407– 414. [10] M. Shilpa et al. , “Enhancing crop yield and growth prediction using IoT-based smart irrigation systems and machine learning algorithms, ” in 2024 Second International Confer ence on Networks, Multimedia and Information T echnology (NMITCON) . IEEE, August 2024, pp. 1–5. [11] D. Doering, A. Benenmann, R. Lerm, E. P . de Freitas, I. Muller , J. M. W inter, and C. E. Pereira, “Design and optimization of a heterogeneous platform for multiple U A V use in precision agriculture applications, ” vol. 47, no. 3, January 2014, pp. 12 272–12 277, 19th IF AC W orld Congress. [12] N. M ¨ uhleis, M. Glaß, L. Zhang, and J. T eich, “ A co-simulation approach for control performance analysis during design space exploration of cyber -physical systems, ” ACM SIGBED Review , vol. 8, no. 2, pp. 23–26, June 2011. [13] J. Fitzgerald, C. Gamble, P . G. Larsen, K. Pierce, and J. W oodcock, “Cyber-physical systems design: Formal foundations, methods and inte- grated tool chains, ” in 2015 IEEE/ACM 3r d FME W orkshop on F ormal Methods in Softwar e Engineering , May 2015, pp. 40–46. [14] M. Herget, F . S. Saadatmand, M. Bor, I. G. Alonso, T . Stefanov , B. Akesson, and A. D. Pimentel, “Design space exploration for dis- tributed cyber-physical systems: State-of-the-art, challenges, and direc- tions, ” in 2022 25th Euromicr o Conference on Digital System Design (DSD) . IEEE, August 2022, pp. 632–640. [15] Y . Liao, T . Adegbija, and R. Lyseck y , “Efficient system-lev el design space exploration for high-level synthesis using Pareto-optimal subspace pruning, ” in Proceedings of the 28th Asia and South P acific Design Automation Confer ence , January 2023, pp. 567–572. [16] Y . Xiao, C. Oh, M. Lora, and P . Nuzzo, “Efficient exploration of cyber -physical system architectures using contracts and subgraph iso- morphism, ” in 2024 Design, Automation & T est in Europe Conference & Exhibition (D ATE) . IEEE, March 2024, pp. 1–6. [17] S.-J. Y u, I. Incer, V . Prabhu, A. Chattoraj, E. Vin, D. Fremont, A. Mehta, A. Sangiovanni-V incentelli, S. Sastry , and S. A. Seshia, “Symbiotic CPS design-space exploration through iterated optimization, ” in Pr oceedings of Cyber-Physical Systems and Internet of Things W eek 2023 , May 2023, pp. 92–99. [18] H. Zheng, J. Betz, A. Ramamurthy , H. Jin, and R. Mangharam, “Com- binatorial and parametric gradient-free optimization for cyber-physical system design, ” in 2022 IEEE W orkshop on Design A utomation for CPS and IoT (DESTION) . IEEE, May 2022, pp. 34–41. [19] Flyability , “Flyability Elios 3, ” 2024, accessed: 2025-02-17. [Online]. A vailable: https://www .flyability .com/hubfs/Kno wledge%20Base% 20Files/Documents/Manuals/E3%20and%20equipments/Elios%203% 20RAD%20User%20Manual%20v1.0.pdf [20] XAG, “XAG V40 HSE-UA V, ” 2024, accessed: 2025-02-17. [Online]. A vailable: https://www .xa.com/en/v40/v40specs [21] Primate Labs, “Geekbench 5 (benchmark), ” 2021, accessed: 2025-9-12. [Online]. A v ailable: https://www .geekbench.com/leg acy/ [22] ——, “Geekbench 5.4.1 (benchmark), ” 2021, accessed: 2025-9-12. [Online]. A v ailable: https://www .primatelabs.com/release/geekbench5/ [23] S. Chetoui and S. Reda, “W orkload- and user-a ware battery lifetime management for mobile SoCs, ” in 2021 Design, Automation & T est in Eur ope Conference & Exhibition (DA TE) . IEEE, 2021, pp. 1679–1684. [24] A. Ignatiev , A. Morgado, and J. Marques-Silva, “PySA T: A Python toolkit for prototyping with SA T oracles, ” in International Conference on Theory and Applications of Satisfiability T esting , June 2018, pp. 428– 437. [25] V . Richthammer, T . Schwarzer, S. W ildermann, J. T eich, and M. Glaß, “ Architecture decomposition in system synthesis of heterogeneous many-core systems, ” in Proceedings of the 55th Annual Design Automa- tion Conference , 2018, pp. 1–6. [26] T . Schwarzer , A. W eichslgartner , M. Glaß, S. Wildermann, P . Brand, and J. T eich, “Symmetry-eliminating design space exploration for hybrid application mapping on many-core architectures, ” IEEE Tr ansactions on Computer-Aided Design of Integrated Circuits and Systems , v ol. 37, no. 2, pp. 297–310, 2017. [27] United States Department of Agriculture, “2022 census of agriculture data, ” 2022, accessed: 2025-06-26. [On- line]. A vailable: https://www .nass.usda.gov/Publications/Highlights/ 2024/Census22 HL FarmsFarmland.pdf [28] Eurostat, “Farms and farmland in the EU, ” 2022, accessed: 2025-09-11. [Online]. A vailable: https://ec.europa.eu/eurostat/statistics- explained/index.php?title=F arms and farmland in the European Union - statistics [29] AllyNav, “ Agriculture robot price - professional farm automation solutions, ” 2025, accessed: 2025-12-18. [Online]. A vailable: https://www .allynav .com/precision- agriculture- solutions/ agricultural- robots/agriculture- robot- price [30] G. Gil et al. , “Why the low adoption of robotics in the farms? challenges for the establishment of commercial agricultural robots, ” Smart Agricultural T echnology , vol. 3, p. 100069, Febuary 2023. [31] United States Department of Agriculture, “National agricul- tural statistics service, ” 2022, accessed: 2025-06-26. [Online]. A vailable: https://www .nass.usda.gov/Statistics by State/W ashington/ Publications/Current News Release/2025/FRUIT .pdf [32] Eurostat, “ Apple orchards in the EU, ” 2019, accessed: 2025-09- 11. [Online]. A vailable: https://ec.europa.eu/eurostat/statistics- explained/ index.php?title=Agricultural production - orchards [33] S. Bloch, “Robotic weeders are racing to re- place glyphosate and dicamba, ” 2019, accessed: 2025-12-18. [Online]. A vailable: https://thecounter .org/ robot- weeders- glyphosate- dicamba- herbicide- replacement/ [34] F . H. Barron, “Selecting a best multiattribute alternativ e with partial information about attribute weights, ” Acta psychologica , vol. 80, no. 1-3, pp. 91–103, 1992. [35] T . V arshney , A. V . W aghmare, V . P . Singh, M. Ramu, N. Patnana, V . P . Meena, A. T . Azar, and I. A. Hameed, “Investig ation of rank order centroid method for optimal generation control, ” Scientific Reports , vol. 14, pp. 1–21, December 2024. [36] E. Roszkowska, “Rank ordering criteria weighting methods – a compar- ativ e overview , ” Optimum. Studia ekonomiczne , pp. 14–33, 2013. [37] B. Suhendarwan, “Rank order centroid as a method to weight dimen- sional variables for de veloping a composite index on policing activities in mining and quarrying areas, ” Journal Researc h of Social Science, Economics, and Management , vol. 3, pp. 1608–1616, March 2024. [38] Duckietown, “Duckiedrone (DD24-B), ” 2025, accessed: 2025- 09-07. [Online]. A vailable: https://get.duckietown.com/products/ autonomous- raspberrypi- quadcopter- duckiedrone- dd24