TS-MAMP: A Remanufactured Agricultural Robot Powered by Second-Life EV Components and NMS-Free On-Device Weed Detection

TS-MAMP: A Remanufactured Agricultural Robot Powered by Second-Life EV Components and NMS-Free On-Device Weed Detection Weijie Shi¹,†, Zicheng Xu²,³,*,†, Zhenbang Cheng¹, Haoran Xuan¹, Mingbo Duan¹, Gan Ge¹ 1 2

School of Mechanical and Automotive Engineering, West Anhui University, Lu'an, 237000, China

School of Automation, Nanjing University of Information Science and Technology, Nanjing, 210000, China 3

School of Engineering, Westlake University, Hangzhou, 310030, China † *

These authors contributed equally to this work.

Corresponding authors: 202542000226@nuist.edu.cn

I.

Abstract—Agriculture 4.0 robotic systems improve field efficiency yet remain too capital-intensive for the fragmented smallholdings that dominate global agriculture. Meanwhile, a growing number of retired low-speed electric-vehicle (LSEV) powertrains retain functional electromechanical value but are destructively recycled. This paper presents TS-MAMP (Telescopic-Sleeve Modular Agricultural Mobile Platform), a remanufactured robot built under 3R (reduce, reuse, recycle) circular-economy principles. Retired 48 V brushless-DC (BLDC) hub motors are paired via back-EMF matching, and lead-acid battery modules screened at 60%–80% state of health are actively balanced within a 100 mV inter-module voltage deviation. Together, these reused components reduce the powertrain-and-chassis BOM cost by approximately 60%, to below USD 450 (perception and weeding modules excluded). The truss chassis provides ≥200 kg static load, continuously adjustable track width from 1200 mm to 2000 mm, and ≤5-minute

module

changeover.

An

NMS-free

(non-maximum-suppression-free) YOLOv10n detector with consistent dual-assignment training and negative-sample learning achieves 80.87% mean average precision (mAP)@0.5 (58.41% mAP@0.5:0.95) on the Wanxi Crop-Weed dataset, and is deployed via FP16 TensorRT on a Jetson Nano, confirming

on-device

inference

feasibility.

TS-MAMP

demonstrates that retired EV components, under modest screening, can be re-engineered into affordable, AI-enabled agricultural robots—opening a remanufacturing pathway for the smallholder fields that commercial automation leaves unserved. Keywords-Agricultural

Many high-throughput Agriculture 4.0 robotic systems are designed for large, standardized farmland and require substantial capital investment. The NEXAT gantry, with its ~1,100 hp drivetrain and 200 t/h harvesting modules, represents a multimillion-pound investment designed for very large-scale commercial farms [1], [2]—illustrating the stark gap between high-end automation and the small, fragmented holdings typical of smallholder agriculture [3], [4]. Modular robots such as Thorvald II have improved configurability across field, greenhouse, and specialty-crop environments [5], [6], while commercial systems like Naïo Technologies' Ted have demonstrated the progress of autonomous mechanical weeding in specialty crops [7]. Yet smallholder adoption remains constrained by capital cost, financial risk, limited credit access, and uncertainty about returns [8]–[10]. Meanwhile, rapid transport electrification is generating a growing stream of retired EV components. In China, retired NEV power batteries are projected to exceed one million tonnes annually by 2030 [11], and comprehensive reviews have confirmed the viability of repurposing these retired batteries for second-life applications

. Additionally,

[12]

electric traction motors contain valuable assemblies that are often lost when end-of-life treatment relies on destructive recycling [13]–[15]. Second-life studies further indicate that batteries retired at approximately 70%–80% SOH retain useful capacity for less demanding stationary or low-speed applications [16]–[19], suggesting an opportunity to connect

robotics;

EV

component

remanufacturing; Circular economy; Edge-deployed weed detection; NMS-free detection

INTRODUCTION

reuse-oriented remanufacturing with low-speed agricultural robotic platforms.

In this work, we present TS-MAMP, a remanufactured

characterization of 20 retired modules and reflects an

agricultural robot built from second-life LSEV components

explicit trade-off between usable residual capacity and

under 3R circular-economy principles

expected

. Unlike

[20]–[23]

cycle

life

under

intermittent,

low-current

high-end gantry or modular platforms that require new

agricultural duty cycles. This approach redirects retired

components and substantial capital, TS-MAMP targets

battery modules from uncontrolled disposal into a controlled

fragmented smallholder terrain through a truss-style,

reuse pathway, reducing the risk of electrolyte leakage and

dimensionally reconfigurable chassis powered by retired

heavy-metal

BLDC motors and lead-acid battery modules, paired with an

end-of-life treatment. The qualified powertrain configuration

NMS-free YOLOv10n perception system

is summarized in Table I.

[24]

that removes

the need for a separate NMS post-processing step on edge hardware. The contributions are threefold: (1) a second-life EV powertrain qualification and modular chassis integration

TABLE I.

contamination

associated

Source (Retired EV) 12 V 20 Ah Graphene-modified Lead-acid Battery × 4

weeding modules excluded); (2) prototype-level mechanical

Energy Storage Traction Motor

verification of ≥200 kg static load, continuously adjustable

Transmission

38-Link Chain & 48T/16T Gear

Control Unit

Brushless EV Controller

Chassis Node

Front Wheel Hub (Model 21-44-20)

by approximately 60% to below USD 450 (perception and

track width from 1200 mm to 2000 mm, and ≤5 min tool-module changeover; and (3) an NMS-free YOLOv10n pipeline

achieving

80.87%

mAP@0.5

(58.41%

improper

CONFIGURATION OF REMANUFACTURED POWERTRAIN AND CHASSIS

Component

methodology reducing the powertrain-and-chassis BOM cost

with

48 V 350 W Hub Brake Motor

Specification in Robot 48 V System (4 Series) 48 V 350 W BLDC Actuator Chain Drive (Reduction Ratio 3:1) Integrated Dual-Mode PWM Driver Differential Steering System

mAP@0.5:0.95) on the Wanxi Crop-Weed dataset, deployed

The fully assembled prototype was built using retired

via FP16 TensorRT on a Jetson Nano for on-device

components, an in-house fabricated steel frame, and

inference. The following sections describe the system design,

off-the-shelf control electronics, with a prototype-level direct

experimental setup, field validation, and future work.

hardware BOM cost below USD 450 covering the traction

II.

powertrain and chassis only (motors, batteries, frame,

SYSTEM DESIGN

controller, battery equalizer, and mechanical drivetrain; the

A. Green Remanufacturing and EV Component Qualification The TS-MAMP platform was designed and fabricated following 3R circular-economy principles

[20]–[23].

The

perception stack — Jetson Nano and IMX219 camera — and the

laser-weeding

module

excluded),

including

transportation, screening labor, and fabrication overhead,

remanufacturing workflow prioritized direct reuse of

approximately

functional electromechanical assemblies salvaged from

new-component baseline comprising commercially sourced

retired low-speed electric vehicles (LSEVs), minimizing

48 V BLDC motors, new lead-acid traction batteries, and a

raw-material extraction and manufacturing energy. Paired

matching frame structure.

DJZ48-10C1 48 V/350 W BLDC hub motors were selected as the traction actuators. Salvaged units were matched by back-EMF waveform and internal resistance prior to installation, preserving kinematic symmetry between the left and right drive units. For energy storage, retired 12 V graphene-modified lead-acid battery modules were collected from LSEV scrapping stations and evaluated with reference to the capacity and safety test methods specified in GB/T 32620.1-2016

[25].

Modules retaining 60%–80% of rated

capacity were selected and reassembled into a 48 V (four-series) / 20 Ah traction pack. The 60%–80% retention window was established through preliminary discharge

60%

lower

than

an

equivalent

B. Mechanical Architecture and Drivetrain The mechanical platform was designed as a twin-parallel bracket truss frame joined by a central support base. The load-bearing

structure

combines

a

high-strength

aluminum-alloy portal frame with Q235 low-carbon-steel beams, providing a compromise between structural rigidity, reduced frame weight, fabricability, and material cost. The rated static load capacity of ≥200 kg was verified through physical prototype loading. The powertrain adopts a chain transmission rather than a belt drive to eliminate slippage under high-dust and humid field conditions. A 3:1 sprocket reduction (16-tooth driving sprocket, 48-tooth driven sprocket) mechanically triples the output torque at the wheel

hubs, creating a torque safety margin that compensates for

vibration-resistant

the expected electromechanical attenuation of second-life

tool-module changeover within 5 minutes.

motors. This mechanical amplification allows the aged BLDC actuators to operate below their peak torque demand during normal traction. The overall mechanical architecture and powertrain layout are illustrated in Fig. 1.

locking

bolts,

allowing

complete

C. Multifunctional Modules and Laser Weeding The TS-MAMP platform utilizes standardized interfaces and rapid connection mechanisms to enable configuration of multifunctional modules. A primary configuration is the laser weeding sub-module, an eco-friendly alternative to chemical weeding [26], [27]. This module employs a 3-DOF Delta parallel manipulator comprising three identical limbs actuated by 42 mm hybrid stepper motors with 0.48 N·m holding torque. A parallelogram guide linkage provides positioning of a four-stage laser system with 4 mm spot diameter and 50 W optical output power. The laser assembly includes a protective enclosure designed with reference to the protective housing requirements of GB/T 7247.1-2024 .

[28]

Figure 1. Mechanical architecture and powertrain components.

To accommodate variable crop row spacing and module reconfiguration, the frame integrates two independent adjustment mechanisms: (i) a continuous telescopic-sleeve system enabling continuously adjustable track width from 1200 mm to 2000 mm, and (ii) an independent sliding-groove system providing lateral module-width adjustment from 300 mm to 600 mm at ±5 mm positioning accuracy.

Both

mechanisms

are

secured

by

(a) Module schematic

Laser activation

is

precisely

controlled

via

the

Modbus-RTU communication protocol and integrated with safety interlock devices. When the system detects that the platform moves at a speed below 0.1 m/s or fails to accurately identify working targets, it activates the laser output suppression mechanism to prevent localized thermal accumulation

during

prolonged

dwell

and

potential

secondary environmental hazards. The multifunctional module configuration is shown in Fig. 2.

(b) Laser weeding submodule structure

Figure 2. Multifunctional module configuration.

D. Electronic Control and Drive System The

electronic

control

system

integrates

power

management, drive control, and onboard perception, enabling

cost-effective

repurposing

of

retired

EV

components for operator-supervised field operation. The powertrain uses salvaged DJZ48-10C1 BLDC hub motors (48 V, 350 W), which are equipped with integrated mechanical expansion brakes, as illustrated in Fig. 3.

(a) DJZ48-10C1 BLDC hub brake motor (b) Workflow schematic Figure 3. Electronic control and drive module.

A dual-channel mixing controller serves as the central electronic interface between the wireless command system

and the drive actuators. It translates 2.4 GHz PWM signals

chinensis) and four associated weeds: Echinochloa crus-galli,

received from a Microzone MC6RE radio receiver into

Eleusine indica, Digitaria sanguinalis, and Portulaca

calibrated analog set-points for two 500 W BLDC motor

oleracea.

speed controllers while maintaining a low-latency manual override via the wireless link for operational safety during system debugging and field maneuvers. To ensure operational safety and alleviate voltage deviation in retired batteries, the system is equipped with an ANGUI KBX104S active battery equalizer. Controlled by an MCU, this device uses a dynamic energy transfer strategy to maintain inter-module voltage differential below 100 mV during both charge and discharge phases. For thermal management, modules are physically separated by 15 mm gaps within a ventilated chassis to facilitate passive air-cooling dissipation. III.

B. NMS-Free Perception Pipeline and Optimization We adopted YOLOv10n with consistent dual assignment (CDA), which supports NMS-free inference through the YOLOv10 dual-assignment

design

. An

[24]

auxiliary

one-to-many head supplies rich supervisory signals for convergence, while a primary one-to-one head assigns exactly one prediction per ground-truth object using a strict matching metric. During inference, only the optimized one-to-one head is retained, removing the need for a separate NMS post-processing step and enabling deterministic edge deployment. Training was performed on an NVIDIA RTX 3090 (24 GB) workstation; inference was deployed on an NVIDIA Jetson Nano via TensorRT FP16 precision, a

DATASET AND EXPERIMENTAL SETUP

deployment route shown to be feasible on such low-power

A. Wanxi Crop-Weed Dataset

edge platforms by recent benchmark studies [29].

To address the domain gap between public crop-weed datasets and local field conditions, we constructed the Wanxi Crop-Weed Dataset in Lu'an, Anhui Province. Images were acquired using the platform's onboard Sony IMX219 CMOS camera at 1920 × 1080 resolution under natural illumination of 20,000–100,000 lux. All data were captured during the 3–5 leaf seedling stage of pak choi, when morphological similarity between crop and associated weeds is most pronounced. The initial dataset comprised 675 images retained at native resolution; for model training, images were resized to 640 × 640 pixels using letterbox padding to preserve aspect ratio. The 675 images were randomly split into training (65%), validation (25%), and test (10%) sets.

To address slow convergence and high false-positive rates in complex seedling-stage scenarios, a two-stage optimization strategy was applied. Stage I expanded the 439-image training partition to 6,020 images via offline dynamic mosaic and photometric augmentation (random horizontal flipping, brightness scaling of ±0.2, and hue shifting of ±0.1). Stage II integrated a hybrid spatial/channel attention mechanism (SAM/CAM) and introduced 500 target-free negative background images, bringing the total training-time sample count to 6,520 with a 7.67% negative-sample ratio. All augmentations and negative samples were applied exclusively to the training partition.

Annotation followed a plant-wise bounding-box protocol

Table II reports the ablation results. The NMS-free

covering five classes: pak choi (Brassica rapa subsp.

YOLOv10n baseline achieved 62.15% mAP@0.5 (45.32% mAP@0.5:0.95).

TABLE II. Optimization Modules Baseline + Data Augmentation + Spatial Attention + Channel Attention + Neg Samples (Final)

QUANTITATIVE ABLATION OF THE NMS-FREE YOLOV10N PERCEPTION PIPELINE.

Precision (%) 72.28 77.03 79.65 81.38 83.31

Recall (%) 67.34 71.29 74.13 75.61 78.41

mAP@0.5 (%) 62.15 69.46 74.22 76.64 80.87

mAP@0.5:0.95 (%) 45.32 49.37 52.81 55.06 58.41

Δ mAP@0.5 (pp) — +7.31 +4.76 +2.42 +4.23

Note: Δ values are step-wise improvements relative to the preceding row.

Introducing data augmentation raised mAP@0.5 to 69.46%

overall improvement of 18.72 percentage points over the

(+7.31 pp); adding spatial attention (SAM) increased it to

baseline. Precision and Recall improved monotonically from

74.22% (+4.76 pp); integrating channel attention (CAM)

72.28%/67.34% to 83.31%/78.41%. Fig. 4 visualizes this

brought it to 76.64% (+2.42 pp); and negative-sample

progression across all four metrics.

learning yielded a final mAP@0.5 of 80.87% (+4.23 pp), an

These results confirm that NMS-free end-to-end detection

can

be

trained

and

deployed

on

resource-constrained edge hardware without sacrificing detection accuracy. Loss curves of both the auxiliary and primary heads showed steady downward trends and rapid stabilization across the baseline and the two optimization Figure 4. Visualization of detection performance across the incremental optimization stages.

stages (Fig. 5). Field detection results from the deployed Jetson Nano platform are shown in Fig. 6.

Figure 5. Training and validation loss comparison across the baseline and the two optimization stages.

A physical prototype of the TS-MAMP was constructed and evaluated in field trials in Lu'an, Anhui (Fig. 7). The tests confirmed the structural integrity of the modular truss chassis and the reliability of the repurposed (second-life) powertrain

under

standard

agricultural

loads.

The

standardized interfaces facilitated rapid switching between functional

modules.

Integrated

with

the

optimized

YOLOv10n perception system, the platform exhibited stable Figure 6. YOLOv10n weed detection results in field conditions.

C. Field Validation

mobility and on-device inference capability, demonstrating the feasibility of repurposing retired EV components for low-cost agricultural automation.

(a) Physical prototype (b) Field test Figure 7. Field validation of the TS-MAMP prototype.

IV.

CONCLUSION

This work demonstrates that repurposed retired EV powertrains can serve as a viable foundation for affordable, intelligent agricultural machinery. By applying 3R principles and screening retired battery modules with reference to GB/T 32620.1-2016, we integrated repurposed 48 V BLDC motors and graphene-modified lead-acid batteries into a functional platform, reducing the powertrain-and-chassis BOM cost by approximately 60% to below USD 450 (perception and weeding modules excluded). On the perception side, the NMS-free YOLOv10n architecture enabled on-device inference on the resource-constrained Jetson Nano, yielding an mAP@0.5 of 80.87% on the Wanxi Crop-Weed dataset. These results establish TS-MAMP as a low-cost pathway that bridges industrial electronic waste recycling with precision farming for smallholder agriculture. The current approach operates within certain boundaries, particularly the capacity heterogeneity of retired battery modules and the absence of multi-modal sensing. Future work will prioritize integrating the active balancing logic into a comprehensive BMS to manage capacity variation, and

enhancing

environmental

robustness

through

LiDAR-vision sensor fusion for all-weather field operation. FUNDING This study was supported by the Anhui Provincial Key Research and Development Project (2024AH051996), the Anhui Provincial Teaching Reform Research Project (2023JYXM0695), and the Anhui AI+Education Curriculum Project (2024AIJY327). REFERENCES [1] [2] [3]

[4]

[5] [6] [7] [8]

T. Chamen and J. McPhee, "Unlocking potentials: Advantages of widespan controlled traffic farming compared to standard CTF," NEXAT GmbH Corporate Technical Note, 2024. O. Mark, "Agritechnica 2023: 1,100hp Nexat gantry system to harvest 200t/hour," Farmers Weekly, Nov. 2023. V. Bulgakov, S. Pascuzzi, V. Adamchuk, V. Kuvachov, and L. Nozdrovicky, "Theoretical study of transverse offsets of wide span tractor working implements and their influence on damage to row crops," Agriculture, vol. 9, no. 7, p. 144, 2019. H. Yang et al., "Achieving the sustainable agricultural development goals by adopting the new energy electric agricultural machinery: An analysis of opportunities and challenges of China," Energies, vol. 18, no. 16, p. 4211, 2025. L. Grimstad and P. J. From, "The Thorvald II agricultural robotic system," Robotics, vol. 6, no. 4, p. 24, 2017. L. Grimstad and P. J. From, "Thorvald II – a modular and re-configurable agricultural robot," IFAC-PapersOnLine, vol. 50, no. 1, pp. 4588–4593, 2017. Naïo Technologies, "Ted – Weeding Robot," 2026. [Online]. Available: https://www.naio-technologies.com/en/ted-robot/ FAO Investment Centre and Innovations for Poverty Action (IPA),

"Pathways to profit: Experimental evidence on agricultural technology adoption," FAO Investment Centre Brief, Oct. 2023. [9] ISF Advisors, "Beyond the frontier: Decoding viability in smallholder finance," 2025 Rural and Agricultural Finance State of the Sector Report, Nov. 2025. [10] T. A. Abetu, P. T. M. Ingenbleek, K. E. Giller, E. Wolde-Meskel, and E. Baars, "Why smallholders (do not) adopt productivity-increasing technologies? Developing a smallholder adoption-process model through an in-depth qualitative study in Ethiopia," Technology in Society, vol. 85, p. 103211, Apr. 2026. [11] Xinhua, "China seeks to step up regulation of NEV battery recycling," May 28, 2026. [12] M. N. Akram and W. Abdul-Kader, "Repurposing second-life EV batteries to advance sustainable development: A comprehensive review," Batteries, vol. 10, no. 12, p. 452, 2024. [13] E. Antwi, G. K. Ayetor, and F. K. Forson, "Review of evaluation and decision-making framework for remanufacturing spent electric vehicle traction motors," Preprints.org, 2026. [14] J. G. Erdmann, "Electromobility: Second life for electric motors," Fraunhofer-Gesellschaft Research News, Jan. 2024. [15] R. Wang, L. Zhan, Z. Xu, R. Wang, and J. Wang, "A green strategy for upcycling utilization of core parts from end-of-life vehicles (ELVs): Pollution source analysis, technology flowchart, technology upgrade," Science of the Total Environment, vol. 912, p. 169609, 2024. [16] M. S. Koroma et al., "Life cycle assessment of battery electric vehicles: Implications of future electricity mix and different battery end-of-life management," Science of the Total Environment, vol. 831, p. 154859, 2022. [17] L. C. Casals, B. A. García, and C. Canal, "Second life batteries lifespan: Rest of useful life and environmental analysis," Journal of Environmental Management, vol. 232, pp. 354–363, 2019. [18] N. Fallah and C. Fitzpatrick, "Exploring the state of health of electric vehicle batteries at end of use; hierarchical waste flow analysis to determine the recycling and reuse potential," Journal of Remanufacturing, vol. 14, no. 1, pp. 155–168, 2024. [19] H. Iqbal, S. Sarwar, D. Kirli, J. K. H. Shek, and A. E. Kiprakis, "A survey of second-life batteries based on techno-economic perspective and applications-based analysis," Carbon Neutrality, vol. 2, no. 1, p. 8, 2023. [20] P. Ghisellini, C. Cialani, and S. Ulgiati, "A review on circular economy: The expected transition to a balanced interplay of environmental and economic systems," Journal of Cleaner Production, vol. 114, pp. 11–32, 2016. [21] V. M. Scharmer et al., "Sustainable manufacturing: A review and framework derivation," Sustainability, vol. 16, p. 119, 2024. [22] D. Tiwari, J. Miscandlon, A. Tiwari, and G. W. Jewell, "A review of circular economy research for electric motors and the role of Industry 4.0 technologies," Sustainability, vol. 13, no. 17, p. 9668, 2021. [23] Z. Li, A. S. Hamidi, Z. Yan, A. Sattar, S. Hazra, J. Soulard, C. Guest, S. H. Ahmed, and F. Tailor, "A circular economy approach for recycling electric motors in the end-of-life vehicles: A literature review," Resources, Conservation and Recycling, vol. 205, p. 107582, 2024. [24] A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding, "YOLOv10: Real-time end-to-end object detection," in Advances in Neural Information Processing Systems 37 (NeurIPS 2024), 2024. [25] GB/T 32620.1-2016, Lead-acid Batteries for Electric Road Vehicles—Part 1: Technical Conditions, AQSIQ/SAC, Beijing, China, 2016. [26] C. Andreasen, K. Scholle, and M. Saberi, "Laser weeding with small autonomous vehicles: Friends or foes?" Frontiers in Agronomy, vol. 4, p. 841086, 2022. [27] P. Zhao, J. Chen, J. Li, J. Ning, Y. Chang, and S. Yang, "Design and testing of an autonomous laser weeding robot for strawberry fields based on DIN-LW-YOLO," Computers and Electronics in Agriculture, vol. 229, p. 109808, 2025. [28] GB/T 7247.1-2024, Safety of Laser Products—Part 1: Equipment Classification and Requirements, State Administration for Market Regulation (SAMR), Beijing, China, 2024. [29] T. P. Swaminathan, C. Silver, and T. Akilan, "Benchmarking deep learning models on NVIDIA Jetson Nano for real-time systems: An empirical investigation," arXiv preprint arXiv:2406.17749, 2024.

Related documents

• Yard and Garden Pest Management#497226
• Radiation detectors and sources enhanced with micro/nanotechnology#958012
• Segmentação de Formigueiros em plantações de cana de açúcar usando imagens de drones#319352
• SCEA-YOLO: A General-Purpose Maturity Grading Model of Multi-Crop Greenhouse Robots.#9225
• Subterranean mine inspection using the Prometheus re-configurable drone#228133