AgriSense: Vision-Language Decision Support and Robotic Intervention in a Virtual Greenhouse Copy title
Organizations: Jeonbuk National University, Deokjin-gu, Jeonju-si, Jeollabuk-do, Republic of Korea · University of Hawai‘i at M¯anoa, Honolulu, Hawaii, United States · Chicago State University, Chicago, Illinois, United States
Abstract
Autonomous greenhouse management increasingly requires robotic systems that integrate perception, navigation, environmental regulation, and plant-level intervention. This paper presents AgriSense, a framework combining an Isaac Sim virtual greenhouse, a quadruped mobile manipulator, RGB and LiDAR-inertial sensing, environmental control, and supervised vision-language decision support. Frozen DINOv2 and BERT encoders, together with structured environmental inputs, support joint prediction of twelve environmental references, a target zone, and an intervention flag. Checkpoint replay reproduces 41/43 zone decisions. Across five seeds on a source-grouped test set, the full, image-removed, and language-removed configurations each achieve 43/44 correct zone decisions, providing no evidence of incremental multimodal zone-accuracy improvement. The full model achieves mean absolute errors of 0.249 degrees C for temperature, 2.030 percentage points for humidity, and 0.997 percentage points for soil moisture. Sensor-only inputs match full-model zone accuracy, while language-only inputs achieve 68.64%. External real-image benchmarks reveal distribution-shift limitations. Five offline Point-LIO routes yield rigidly aligned trajectory errors of 5.22-17.32 mm. Of 40 virtual-arm targets, 32 satisfy geometric acceptance criteria and eight are rejected at boundaries. Six long-horizon climate cases and fault protocols characterize integration limits. Navigation and integrated treatment rely on simulator-state inputs; physical greenhouse performance and agronomic treatment efficacy remain unvalidated. The contribution is an evaluated decision-support and robotic-intervention framework for a virtual greenhouse.
Figures & tables
| Study | Primary task/output | Relation to the present study |
| AgriVLN ( Zhao et al., 2025 ) | Instruction-following agricultural navigation | Evaluates navigation rather than environmental-reference, target-zone, and intervention prediction. |
| Horticultural planning ( Cuaran et al., 2026 ) | Visual-language planning over robot primitives | A planner and a fixed-format predictor have different evaluation targets. |
| AgroGPT ( Awais et al., 2025 ) | Conversational agricultural visual interpretation | Evaluates agricultural visual question answering rather than the reference/zone/intervention outputs studied here. |
| AGRI-BT ( Salman et al., 2026 ) | Language-generated behaviour trees and visual perception | Represents task-planning integration rather than this supervised fusion model. |
| AgriSense | Image–text–environment prediction plus separate robotic/control interfaces | Evaluated in a virtual greenhouse; no claim of new encoders, a new LIO algorithm, or field-validated treatment. |
| Module | Inputs | Outputs | Representative interface or artifact |
| Robot sensing | RGB, LiDAR, IMU, proprioception | Images, point clouds, inertial measurements | Robot-view image stream, LiDAR cloud, IMU and robot-state topics |
| Zone sensing | Fixed environmental sensor nodes | Temperature, humidity, and soil-moisture states | Global and per-zone ROS 2 messages and synchronized CSV logs |
| LiDAR–IMU spatial grounding | Point clouds, IMU, local map | Robot pose, trajectory, current zone, map update | Local odometry, reconstructed point cloud, zone-association output |
| Multimodal dataset recorder | Image, prompt, zone, pose, and environmental state | Aligned multimodal samples and target labels | Dataset CSV, image directory, preprocessing metadata |
| VLDA inference | Prompt, RGB image, environmental vector, current zone | Setpoint proposals, target zone, spray-intervention probability | Prediction packet and offline evaluation outputs |
| Validation and command dispatch | Model outputs, timestamps, pose, map, actuator limits | Accepted command, rejected proposal, or fallback request | Validated command buffer, event log, execution-mode state |
| Subset | Base/edited | Total | Target zones 1/2/3/4 | Negative/positive |
| Historical training | 105/93 | 198 | 27/26/125/20 | 105/93 |
| Historical validation | 22/20 | 42 | 6/6/26/4 | 22/20 |
| Historical test | 23/20 | 43 | 6/5/28/4 | 23/20 |
| All supervised rows | 150/133 | 283 | 39/37/179/28 | 150/133 |
| Grouped split | 197 training / 42 validation / 44 test rows; 265 source/similarity components | |||
| Item | Historical checkpoint | Repeated branch evaluation |
| Backbones/pooling | Frozen DINOv2-small and BERT; CLS/CLS | Same frozen weights; cached evaluation-mode CLS features |
| Trainable capacity | 514,065 parameters | Same stored projection/encoder/fusion/head topology; removed features zeroed |
| Optimizer | AdamW; learning rate , weight decay | AdamW; same rates |
| Batch/epoch policy | 16; maximum 300; patience 10 | 16; maximum 100; patience 10 |
| Recorded history | 50 epochs; selected epoch 40 | Validation-selected per run; per-run records supplied |
| Seeds | 42 | 17, 29, 43, 71, 101 |
| Field group | Contents |
| Image | Robot-view RGB image or controlled augmented crop-view sample |
| Language | User request, mission prompt, or task instruction |
| Sensor and zone context | Current-zone code and global/per-zone temperature, humidity, and soil-moisture measurements |
| Regression targets | Twelve environmental references: temperature, humidity, and soil moisture for Zones 1–4 |
| Classification target | Target management zone |
| Binary-action target | Apply or skip localized pesticide intervention |
| Metric | Original evaluation | Checkpoint replay |
| Target-zone accuracy | 95.35% (41/43) | 95.35% (41/43) |
| Target-zone macro-F1 | 0.9311 | 0.9311 |
| Intervention-label agreement | 100% (43/43) | 100% (43/43) |
| Mixed-unit reference MAE | 1.266192 | 1.265588 |
| Temperature MAE ( ∘ C) | Not separated | 0.323820 |
| Humidity MAE (percentage points) | Not separated | 2.234821 |
| Condition | Zone accuracy (%) | Macro-F1 | Temp. MAE ( ∘ C) | RH MAE (pp) | Soil MAE (pp) |
| Full model | |||||
| Without language | |||||
| Without image | |||||
| Without sensor/current zone | |||||
| Sensor/current-zone only | |||||
| Language only |
| Model/regime | Runs | PV accuracy (%) | External accuracy (%) | External macro-F1 |
| ResNet-50 / frozen probe | 5 | |||
| EfficientNet-B0 / frozen probe | 5 | |||
| ViT-B/16 / frozen probe | 5 | |||
| CLIP / image probe | 5 | |||
| CLIP / fixed zero-shot | 1 | 12.566 | 30.252 | 0.2546 |
| ResNet-50 / full network | 3 |
| Route | Length (m) | Initial ATE (mm) | Rigid-fit ATE (mm) | RPE 1 s (mm) |
| 00 | 26.89 | 19.80 | 5.22 | 3.80 |
| 01 | 26.83 | 40.12 | 9.63 | 4.77 |
| 02 | 26.97 | 74.26 | 17.32 | 6.26 |
| 03 | 27.05 | 20.37 | 5.71 | 4.16 |
| 04 | 27.12 | 22.89 | 6.35 | 4.36 |
| Case/quantity | Reported outcome | Interpretation |
| High temperature | Within 0.5 ∘ C after 2,461 s | Sustained settling reported for this case |
| High humidity | Within 2 percentage points after 3,448 s | Sustained settling reported for this case |
| Dry-substrate VWC | 90% recovery after 7,457 s; final 0.33266 | Does not meet sustained VWC settling |
| Hot/dry | Final 42.97 ∘ C and 10.64% RH | Temperature, RH, and VWC fail settling |
| Combined disturbance | VWC fails sustained settling | Multi-disturbance operating limit |
| Actuator saturation | Temperature fails sustained settling | Actuator-limited operating envelope |
| Case | Completion time (s) | Maximum treated-target ray error (mm) |
| Four-stop full route | 130.302 | 12.94 |
| Left berry | 30.802 | 10.99 |
| Right berry | 33.802 | 6.40 |
| Component | Quantitative evidence | Boundary |
| VLDA | Original 41/43; grouped full/no-image/no-text/sensor-only 43/44 per seed; language-only 68.64%; per-variable MAE | One episode and generated labels; no demonstrated zone synergy |
| Visual baselines | 20 PV probes; nine full-network runs; external and corruption results | Different target space from VLDA; poor compatible transfer |
| Point-LIO | Five actual offline reconstructions; ATE and 1 s RPE | Synthetic streams; not navigation feedback |
| Navigation/control | 10 nominal, 6 clear, 2 obstacle, 20 valid controls | Arm-proxy protocol; simulator pose |
| Manipulator | 32 accepted targets; eight boundary rejects; three missions | Geometry/oracle targets, not deposition or crop response |
| Environmental model | Six 12-hour cases; 6/18 settling failures | Synthetic weather; incomplete process parameterization |
Appendix figures & tables59 assets
Supplementary material from the paper’s appendix.
Appendix
| Component | Implemented operation | State |
| Image encoder | DINOv2-small CLS 384; projection 384 to 256 | Backbone frozen; projection trainable |
| Text encoder | BERT-base CLS 768; max 128 tokens; projection 768 to 128 | Backbone frozen; projection trainable |
| Sensor/context branch | 19 inputs: current-zone one-hot (4) + continuous context (15); MLP 19 to 128 to 128 | Trainable; training-fitted standardization |
| Fusion | Concatenate 256+128+128=512; MLP 512 to 256 to 256 | Trainable; concatenation, not cross-modal attention |
| Three heads | Each 256 to 128, then 12 setpoints / 4 zone logits / 1 pesticide logit | Trainable; no pixel-mask or trajectory output |
| Objective | MSE + class-weighted CE + positive-weighted BCE-with-logits | Outer loss coefficients all 1 |
| Base row | Aug. row | Context | 12 targets | Prompt | Zone | Pesticide |
|---|---|---|---|---|---|---|
| 102 | 152 | Identical | Identical | Different | 3 to 2 | False to True |
| 130 | 163 | Identical | Identical | Different | 3 to 2 | False to True |
| 122 | 164 | Identical | Identical | Different | 3 to 2 | False to True |
| 108 | 170 | Identical | Identical | Different | 3 to 2 | False to True |
| 26 | 181 | Identical | Identical | Different | 3 to 1 | False to True |
| 131 | 252 | Identical | Identical | Different | 3 to 2 | False to True |
| Data or derived evaluation | Samples | Train / validation / test | Interpretation |
| Original episode, historical split | 283 | 198 / 42 / 43 | 150 base + 133 disease-named augmentations; teacher/simulation labels |
| Same original episode, grouped split | 283 | 197 / 42 / 44 | 265 screened source/similarity components; one episode |
| PlantVillage retained v2 | 49,098 | 34,360 / 4,910 / 9,828 | 54,305 decode; 5,207 quarantined |
| PlantSeg eligible | 7,551 | 5,228 / 821 / 1,502 | 7,774 image/mask pairs decode; 223 unique exclusions |
| Compatible external test | 238 | – / – / 238 | 15 true classes; 238 groups; all 38 outputs retained |
| Novel-host OOD test | 671 | – / – / 671 | 19 literal host labels; 666 groups |
| Model regime | Trainable params | Extract s | Train s | Peak MiB | Latency ms |
|---|---|---|---|---|---|
| ResNet-50 frozen | 77862 | 25.59 | 1.60 | 471.2 | 2.800 |
| EffNet-B0 frozen | 48678 | 13.72 | 1.52 | 437.9 | 3.863 |
| ViT-B/16 frozen | 29222 | 49.96 | 2.26 | 799.5 | 5.130 |
| CLIP probe frozen | 19494 | 44.49 | 3.52 | 773.4 | 9.763 |
| ResNet-50 full | 23585894 | – | 427.95 | 3322.5 | 2.815 |
| EffNet-B0 full | 4056226 | – | 272.86 | 3268.3 | 3.688 |
| Metric | Point | Observed-group interval | Valid draws |
| Zone accuracy (41/43), % | 95.349 | 1000/1000 | |
| Zone macro-F1, fixed four classes | 0.9311 | 1000/1000 | |
| Zone balanced accuracy | 0.9494 | 971/1000 | |
| Pesticide generated-label agreement (43/43) | 1.0000 | 1000/1000 |
| Condition | Zone accuracy % | Four-class macro-F1 | Temperature MAE ∘ C | RH MAE pp | Soil MAE pp |
| Full | |||||
| Without language | |||||
| Without sensor/current-zone | |||||
| Without image | |||||
| Sensor/current-zone only | |||||
| Language only |
| Cohort / model | Runs | Accuracy % SD pp | Accuracy interval % | Macro-F1 SD |
| Frozen probe / ResNet-50 | 5 | |||
| Frozen probe / EfficientNet-B0 | 5 | |||
| Frozen probe / ViT-B/16 | 5 | |||
| Frozen probe / CLIP image probe | 5 | |||
| CLIP image–text zero-shot | 1 | |||
| Full network / ResNet-50 | 3 |
| Cohort / model / condition | Runs / | Acc. | Macro-F1 | Bal. acc. | ECE |
|---|---|---|---|---|---|
| External frozen / CLIP probe / external clean | 5 / 238 | 24.87 0.69 | 0.087 0.003 | 0.250 0.005 | 0.532 0.014 |
| External frozen / CLIP zero-shot / external clean | 1 / 238 | 30.25 | 0.100 | 0.240 | 0.091 |
| External frozen / EffNet-B0 / external clean | 5 / 238 | 24.62 0.70 | 0.091 0.004 | 0.262 0.016 | 0.416 0.013 |
| External frozen / ResNet-50 / external clean | 5 / 238 | 20.17 1.19 | 0.070 0.003 | 0.203 0.010 | 0.567 0.016 |
| External frozen / ViT-B/16 / external clean | 5 / 238 | 21.68 0.97 | 0.064 0.005 | 0.193 0.009 | 0.500 0.007 |
| External full / EffNet-B0 / external clean | 3 / 238 | 25.77 0.87 | 0.085 0.005 | 0.252 0.013 | 0.355 0.007 |
| True class | Full R50 | Full EB0 | Full ViT | Frozen ViT | CLIP zero | |
|---|---|---|---|---|---|---|
| Apple: Apple scab | 123 | 0.997 | 0.999 | 0.997 | 0.985 | 0.000 |
| Apple: Black rot | 124 | 1.000 | 1.000 | 0.996 | 0.996 | 0.000 |
| Apple: Cedar apple rust | 56 | 0.994 | 1.000 | 0.994 | 1.000 | 0.000 |
| Apple: healthy | 308 | 0.993 | 0.997 | 0.999 | 0.993 | 0.000 |
| Blueberry: healthy | 279 | 0.997 | 0.999 | 1.000 | 0.998 | 0.000 |
| Cherry (including sour): Powdery mildew | 206 | 0.998 | 0.997 | 1.000 | 0.995 | 0.000 |
| Cohort / model | Brightness | Blur radius 2 | Noise |
| Frozen probe / ResNet-50 | |||
| Frozen probe / EfficientNet-B0 | |||
| Frozen probe / ViT-B/16 | |||
| Frozen probe / CLIP image probe | |||
| CLIP image–text zero-shot | |||
| Full network / ResNet-50 |
| Cohort / model | Runs | PV accuracy % | External accuracy SD pp | External F1 15 SD | External–PV gap pp [interval] |
| Frozen probe / CLIP image probe | 5 | ||||
| Frozen probe / EfficientNet-B0 | 5 | ||||
| Frozen probe / ResNet-50 | 5 | ||||
| Frozen probe / ViT-B/16 | 5 | ||||
| Full network / EfficientNet-B0 | 3 | ||||
| Full network / ResNet-50 | 3 |
| Cohort / model | Runs | OOD AUROC SD | OOD AP SD | ID rejected % SD pp | OOD accepted % SD pp |
| Frozen probe / ResNet-50 | 5 | ||||
| Frozen probe / EfficientNet-B0 | 5 | ||||
| Frozen probe / ViT-B/16 | 5 | ||||
| Frozen probe / CLIP image probe | 5 | ||||
| Full network / ResNet-50 | 3 | ||||
| Full network / EfficientNet-B0 | 3 |
| Cohort / model | Runs | Global-ID AUROC SD | Restricted-ID AUROC SD | OOD accepted % SD pp |
| CLIP fixed image–text zero-shot | 1 | |||
| Frozen probe / CLIP image probe | 5 | |||
| Frozen probe / EfficientNet-B0 | 5 | |||
| Frozen probe / ResNet-50 | 5 | |||
| Frozen probe / ViT-B/16 | 5 | |||
| Full network / EfficientNet-B0 | 3 |
| Model regime | ID reference | Runs | AP | ID reject % | OOD accept % |
|---|---|---|---|---|---|
| Fixed zero / CLIP probe | Global | 1 | 0.0096 | 5.91 | 98.47 |
| Fixed zero / CLIP probe | Host-restricted | 1 | 0.1155 | 10.59 | 98.47 |
| Frozen / CLIP probe | Global | 5 | 0.2226 0.0138 | 6.07 0.17 | 37.86 3.44 |
| Frozen / CLIP probe | Host-restricted | 5 | 0.7693 0.0121 | 3.92 0.26 | 37.86 3.44 |
| Frozen / EffNet-B0 | Global | 5 | 0.1982 0.0254 | 5.17 0.05 | 37.86 1.58 |
| Frozen / EffNet-B0 | Host-restricted | 5 | 0.7876 0.0083 | 3.37 0.26 | 37.86 1.58 |
| Protocol | Metric | Reported value |
|---|---|---|
| arm-proxy nominal | completion | 10/10 cases |
| arm-proxy nominal | reported Wilson95 interval | 72.2–100.0 percent |
| arm-proxy clear robustness | completion | 6/6 cases |
| arm-proxy physical obstacles | intended safe-stop outcomes | 2/2 cases |
| arm-proxy valid controls | completion | 20/20 cases |
| navigation command | aisle-transit speed command | 0.50 m/s |
| Protocol | Metric | Reported value |
|---|---|---|
| Capacitive soil VWC probes | authored component count | 24 components |
| Shielded temperature/RH stations | authored component count | 4 components |
| Zone irrigation valves | authored component count | 4 components |
| Exhaust fans | authored component count | 4 components |
| Suspended circulation fans | authored component count | 3 components |
| Protocol | Metric | Reported value |
|---|---|---|
| virtual crop-edge targets | pre-specified total requests | 40 requests |
| accepted virtual targets | position error mean sampleSD | 0.449 0.154 mm |
| accepted virtual targets | maximum position error | 0.853 mm |
| accepted virtual targets | mean nozzle-axis error | 0.000502 rad |
| accepted virtual targets | pose acceptance | 32/32 requests |
| accepted virtual targets | virtual disk hits | 32/32 requests |
| Case / seed | Duration s | Max treated-target ray error mm |
|---|---|---|
| full route / 1200 | 130.302 | 12.94 |
| left berry / 1201 | 30.802 | 10.99 |
| right berry / 1202 | 33.802 | 6.40 |
| Route | Length m | Initial ATE mm | Fit ATE mm | 1 s RPE mm |
|---|---|---|---|---|
| route_00 | 26.89 | 19.80 | 5.22 | 3.80 |
| route_01 | 26.83 | 40.12 | 9.63 | 4.77 |
| route_02 | 26.97 | 74.26 | 17.32 | 6.26 |
| route_03 | 27.05 | 20.37 | 5.71 | 4.16 |
| route_04 | 27.12 | 22.89 | 6.35 | 4.36 |
| Protocol | Metric | Reported value |
|---|---|---|
| environmental suite | Cases, duration and logging rate | 6 cases; 12 h each; 1 Hz logging |
| high temperature | Sustained settling within 0.5 degrees C | 2461 s |
| high humidity | Sustained settling within 2 RH percentage points | 3448 s |
| dry substrate | 90% recovery time | 7457 s |
| dry substrate | final VWC | 0.33266 fraction |
| all variable/case rows | sustained settling failures | 6/18 rows |
| Fixture or evidence | Observed outcome | Permitted interpretation |
| Engineered missing/stale/desynchronized inputs | 52 rejected / 52 invalid fixtures | Existing presence/age/synchrony branches match specification |
| Fresh valid control | 1 accepted / 1 valid control; false blocks 0/1 | Only one valid control; no population false-block rate |
| Nonfinite/extreme numeric inputs | 64 accepted / 64 probes | This method does not enforce finite/range bounds |
| Synchronized future timestamp | 1 accepted / 1 probe | Future-time admissibility not enforced by this method |
| Expert intervention outcomes | Unavailable; no unsafe-action/false-intervention rate | Generated-label agreement is not field safety |
| File under supplement_data/ | Contents and interpretation |
|---|---|
| classification_per_run_230.csv | All 230 portable classification evaluations, metrics and conditional CI valid counts; not 230 trainings. |
| classification_model_means_56.csv | 56 task/condition/model families, seed means and sample SD. |
| classification_interval_rows_224.csv | Four aggregate interval metrics per family, with finite/requested counts and scope. |
| classification_all_per_class_230.csv | All class precision/recall/F1, confusion counts, applicable ROC/AP and calibration rows with run identity. |
| classification_curve_and_draw_catalog_230.csv | Relative paths, sizes and hashes for full per-run curves, confusion, calibration, risk–coverage and draws in the validation package. |
| matched_taxonomy_transfer.csv | Per-seed and aggregate post hoc shared-15-class transfer contrasts; all 38 outputs retained. |
| Metric | Value |
| Number of test samples | 43 |
| Overall setpoint MAE | 1.266 |
| Target-zone accuracy | 95.35% |
| Target-zone macro-F1 | 93.11% |
| Spray-label accuracy (exploratory) | 100.00% |
| Spray-label F1 (exploratory) | 100.00% |
| Mission | Pairs | Time [s] | Path [m] | XY RMSE [cm] | XY P95 [cm] | XY Max [m] | 3D RMSE [cm] | Sync [ms] |
| Inspect All Zones | 7643 | 607.95 | 151.23 | 5.34 | 9.94 | 0.18 | 5.64 | 69.93 |
| Inspect Zone 2 | 2893 | 235.54 | 50.21 | 3.43 | 7.60 | 0.10 | 3.69 | 80.78 |
| Inspect Zone 4 | 3181 | 222.94 | 81.34 | 5.16 | 8.98 | 0.11 | 5.39 | 80.38 |
| Mission | Variable | Zone 1 | Zone 2 | Zone 3 | Zone 4 | Min | Max |
| Inspect All Zones | Temperature [ ∘ C] | 0.00 | -0.00 | 1.01 | -1.40 | 20.27 | 25.38 |
| Inspect All Zones | Humidity [%] | 12.70 | 17.01 | 30.99 | 11.65 | 40.68 | 81.00 |
| Inspect All Zones | Soil moisture [%] | 0.02 | -25.38 | -12.48 | -13.17 | 62.00 | 92.18 |
| Inspect Zone 2 | Temperature [ ∘ C] | -2.87 | -1.69 | 1.25 | -0.31 | 20.28 | 25.37 |
| Inspect Zone 2 | Humidity [%] | 11.12 | 5.79 | 8.79 | -1.84 | 58.00 | 81.00 |
| Inspect Zone 2 | Soil moisture [%] | -22.20 | -14.72 | -9.58 | -21.23 | 62.00 | 92.18 |
| Criterion | Zone-level regulation | Plant-level intervention |
| Spatial scope | Entire management zone | Selected plant or local region |
| Primary actuator | Fixed greenhouse actuator | Mobile manipulator and spray end-effector |
| Spatial requirement | Zone association | Zone, inspection pose, and local target |
| Decision context | Environmental measurements and references | Image, state, pose, and reachability |
| Available evidence | Recorded zone states and proposed references | Qualitative simulated command execution |
| Main uncertainty | Environmental dynamics and controller tuning | Target localization, reachability, and spray deposition |
| Mission | Directory and supplied artifacts |
| All zones | inspect_allzones/ : synced_errors.csv , greenhouse_state.csv , metrics_summary.csv , and live_dashboard.html |
| Zone 2 | inspect_zone2/ : synced_errors.csv , greenhouse_state.csv , metrics_summary.csv , and live_dashboard.html |
| Zone 4 | inspect_zone4/ : synced_errors.csv , greenhouse_state.csv , metrics_summary.csv , and live_dashboard.html |
| Field | Meaning | Representative values | Evaluation output |
| Mission prompt | High-level operator request | Inspect Zone 2; inspect all zones; regulate dry soil; inspect and treat an affected crop | Prompt interpretation and task proposal |
| Initial robot pose | Starting pose in the local greenhouse frame | Entrance, central aisle, or zone-specific starting location | Localization initialization and mission repeatability |
| Target zone | Requested or inferred management zone | Zones 1–4 or complete-greenhouse inspection | Target-zone prediction and navigation-goal consistency |
| Environmental condition | Global and per-zone sensor state | High humidity, low soil moisture, elevated temperature, or mixed disturbances | Environmental-reference prediction and PI/PID response |
| Visual condition | Crop and scene appearance in the RGB input | Healthy, disease-like, occluded, low-light, or viewpoint-varied | Inspection priority and intervention proposal |
| Localization mode | Spatial-grounding configuration | GNSS denied; LiDAR–IMU localization and local mapping | Trajectory error, map consistency, and zone association |
Agricultural Robotics
Agricultural robotics develops autonomous machines that sense, navigate, and perform tasks such as monitoring, weeding, spraying, and harvesting crops.
See all 101 papers