HortiMulti: A Multi-Sensor Dataset for Localisation and Mapping in Horticultural Polytunnels
Authors: Shuoyuan Xu, Zhipeng Zhong, Tiago Barros, Matthew Coombes, Cristiano Premebida, Hao Wu, Cunjia Liu
Organizations: Department of Aeronautical and Automotive Engineering, Loughborough University, Leicestershire, UK, LE11 3TU · Department of Electrical and Computer Engineering, Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal · Antobot Ltd, Arise Innovation Hub, Chelmsford, Essex, UK, CM1 1SQ
Agricultural robotics is gaining increasing relevance in both research and real-world deployment. As these systems are expected to operate autonomously in more complex tasks, the availability of representative real-world datasets becomes essential. While domains such as urban and forestry robotics benefit from large and established benchmarks, horticultural environments remain comparatively under-explored despite the economic significance of this sector. To address this gap, we present HortiMulti, a multimodal, cross-season dataset collected in commercial strawberry and raspberry polytunnels across an entire growing season, capturing substantial appearance variation, dynamic foliage, specular reflections from plastic covers, severe perceptual aliasing, and GNSS-unreliable conditions, all of which directly degrade existing localisation and perception algorithms. The sensor suite includes two 3D LiDARs, four RGB cameras, an IMU, GNSS, and wheel odometry. Ground truth trajectories are derived from a combination of Total Station surveying, AprilTag fiducial markers, and LiDAR-inertial odometry, spanning dense, sparse, and marker-free coverage to support evaluation under both controlled and realistic conditions. We release time-synchronised raw measurements, calibration files, reference trajectories, and baseline benchmarks for visual, LiDAR, and multi-sensor SLAM, with results confirming that current state-of-the-art methods remain inadequate for reliable polytunnel deployment, establishing HortiMulti as a one-stop resource for developing and testing robotic perception systems in horticulture environments.
Figures & tables
Figure 1: HortiMulti is a multi-seasonal dataset collected in strawberry (left) and raspberry (right) polytunnels. Data are synchronised across multiple LiDARs and multiple cameras; additional sensors include an AHRS, GNSS, and wheel odometry (bottom). Ground truth is obtained using AprilTags and surveying equipment (bottom right).
Dataset
Platform
Lighting
Season
Sensor
Ground Truth
LiDAR
Mono
Stereo
IMU
Wheel Odom
Forest and Off-Road
TB-Places [ 16 ]
UGV
x
x
x
✓
x
✓
✓
Laser tracker
Wild-Places [ 17 ]
Handheld
x
✓
✓
✓
x
✓
x
SLAM
FinnForest [ 18 ]
Car
✓
✓
x
✓
✓
✓
x
GNSS/INS
SFU Mountain [ 19 ]
UGV
✓
x
✓
✓
✓
✓
✓
Manually-aligned locations
Table 1: Summry of agricultural / off-road SLAM datasets.
Figure 2: Robot platform and sensor tower
Symbol
Dim. (mm)
Symbol
Dim. (mm)
Symbol
Dim. (mm)
A
48
G
121
M
303
B
51
H
199
N
160
C
210
I
43
O
113
D
573
J
197
P
108
E
330
K
505
Q
63
F
340
L
305
R
76
Table 2: Geometric parameters of the mobile sensing platform (labels refer to Fig. 2 )
Sensor
Model
Specification
Spinning LiDAR
Velodyne VLP-16
16 channels, 360 ∘ × 40 ∘ FOV, 200 m range, 10 Hz, ±3 cm accuracy
Spinning LiDAR
Ouster OS0-64
64 channels, 360 ∘ × 90 ∘ FOV, 100 m range, 10 Hz, ±0.8 cm accuracy
Figure 3: Example of a colourised point cloud with corresponding image (bottom left).
Figure 4: Example of some representative trajectories, designed to mimic real farm machinery operations: cross-tunnel passes that mimic spraying tasks, and row-wise traverses within a single polytunnel that resemble UV treatment and potential harvesting robots.
Sensor
Topic Name
ROS Message Type
Rate
Ouster LiDAR
/ouster/points
sensor_msgs/PointCloud2
∼ 10 Hz
Velodyne LiDAR
/velodyne_points
sensor_msgs/PointCloud2
∼ 10 Hz
Forward-Left Camera
/forwardLeft/image_raw/compressed
sensor_msgs/CompressedImage
∼ 10 Hz
/forwardLeft/camera_info
sensor_msgs/CameraInfo
∼ 10 Hz
/forwardLeft/tag_detections
apriltag_ros/AprilTagDetectionArray
∼ 10 Hz
Forward-Right Camera
/forwardRight/image_raw/compressed
sensor_msgs/CompressedImage
∼ 10 Hz
Table 5: Topics stored in each data session.
Figure 6: Ground truth data collection: (a) and (b) illustrate AprilTag installations in strawberry and raspberry polytunnels, respectively; (c)–(f) show the total station and prism setup used to establish a control network and survey the AprilTag positions; (g)–(i) present representative survey results, with (g) showing all tags within a polytunnel, (h) showing all tags without polytunnel obstruction, and (i) indicating how many tags in (h) were surveyed.
Sequence
Frames
Row Count
Query Pos.
Valid Loops
Loop Ratio (%)
Min Dist (m)
Mean Dist (m)
Median Dist (m)
Max Dist (m)
Std Dist (m)
Nov2025/Strawberry_1
932
4
437
437
47
0.008
0.220
0.214
0.687
0.110
Jun2025/Strawberry_1
1953
7
723
723
37
0.007
0.936
0.244
9.850
2.096
Jun2025/Raspberry_1
340
4
181
181
53
0.017
0.225
0.168
1.929
0.267
Aug2025/Raspberry_1
697
8
394
394
57
0.013
1.074
0.349
7.898
1.827
Table 6: Ground truth statistics for place recognition. Configuration used for all sequences: warm-up frames (skip first N) = 100, lower-bound frame gap (ignore immediate past) = 50, distance threshold = 10 m, and top-1 closest neighbour.
Configuration
RMSE (m)
ME (m)
Max error (m)
Poly-TagSLAM
0.1414
0.1269
0.2937
Poly-TagSLAM + Total Station
0.0726
0.0351
0.2633
LOO validation (held-out)
0.0655
0.0586
0.1403
Table 7: Landmark Mapping Accuracy
Figure 7: Cumulative distribution of localisation errors for held-out landmarks under LOO cross-validation.
Figure 8: Qualitative trajectory comparisons within the polytunnel. (a, b) sequence Feb2026 / Strawberry _1 , (c, d) sequence Feb2026 / Strawberry _2 (e, f) sequence Feb2026 / Strawberry _3 , and (g, h) sequence Feb2026 / Strawberry _4 . These figures illustrate the performance and stability of SLAM algorithms with different designed robot trajectories, showing variations in drift, trajectory consistency, and overall accuracy
Stat/Sequence (Feb2026/Strawberry_)
1
2
3
4
RPE (%)
ATE (m)
RPE (%)
ATE (m)
RPE (%)
ATE (m)
RPE (%)
ATE (m)
Duration (s)
876
952
242
348
Distance (m)
951.60
940.90
231.02
350.37
Size (GB)
37.1
40.3
10.2
16.9
DSO (Mono)
×
×
32.07
10.2568
43.53
14.1276
22.84
2.2490
ORB3 (Mono)
×
×
×
×
19.42
0.4920
11.16
1.5589
Table 8: Sequences specifications and SOTA SLAM performance comparison
Figure 9: Retrieval performance for the top-25 candidates within a 10 m range
Jun2025/Strawberry_1
Nov2025/Strawberry_1
Jun2025/Raspberry_1
Aug2025/Raspberry_1
@5m
@10m
@25m
@5m
@10m
@25m
@5m
@10m
@25m
@5m
@10m
@25m
SPVSoAP3D [ 51 ]
0.201
0.245
0.356
0.251
0.306
0.384
0.633
0.689
0.768
0.309
0.431
0.519
PointNetVLAD [ 52 ]
0.452
0.508
0.598
0.531
0.591
0.660
0.836
0.876
0.944
0.212
0.366
0.428
PointNetPGAP [ 53 ]
0.504
0.576
0.668
0.508
0.637
0.692
0.797
0.842
0.915
0.430
0.558
0.596
LoGG3D-Net [ 54 ]
0.068
0.094
0.148
0.083
0.110
0.182
0.282
0.356
0.554
0.091
0.177
0.265
OverlapTransformer [ 55 ]
0.028
0.045
0.076
0.094
0.136
0.209
0.254
0.322
0.395
0.064
0.171
0.227
Table 9: Retrieval performance, using recall, of the top-1 retrieved candidate for various threshold range.
Figure 10: Illustration of the top-1 true positive predictions within a 10 m range. For clear visualization only a fraction of the true positives are plotted.
Robotic strawberry harvesting requires precise 6D pose estimation; however, collecting 6D pose ground truth in real agricultural fields is inherently challenging. Existing strawberry 6D pose estimation studies have therefore relied mainly on synthetic data, leaving their in-field performance unquantified. In this work, we obtain ground truth indirectly, by recovering camera poses via PnP, reconstructing each scene at metric scale, and annotating a single 3D bounding box per strawberry that is propagated across all frames, yielding, to the best of our knowledge, the first real-world 6D pose ground-truth dataset of red-stage strawberries collected at an actual strawberry farm (12,040 images). We also introduce a synthetic dataset rendered in NVIDIA Isaac Sim, featuring scene-level realism and domain randomization. Despite this improved simulation setup, models trained on synthetic data alone fail to transfer to in-field images, while introducing a small amount of real data improves both translation and rotation accuracy across all backbones. Under the monocular RGB setting evaluated here, rotation is largely recovered once real data is used, and depth is what limits pose accuracy. These baselines across backbone encoders serve as a reference for future work. The real-world dataset is publicly available at https://huggingface.co/datasets/WoojungSon/FieldStraw6D, and the data-collection pipeline is available at https://github.com/wjson2435/FieldStraw6D-pipeline.
Woojung Son, Won Suk Lee, Zijing Huang +4
Department of Agricultural and Biological Engineering, University of Florida · Department of Electrical and Computer Engineering, University of Florida · Edwardson School of Industrial Engineering, Purdue University +1
Reliable long-term deployment of autonomous robots in agricultural environments remains challenging due to perceptual aliasing, seasonal variability, and the dynamic nature of crop canopies. Vineyards, characterized by repetitive row structures and significant visual changes across phenological stages, represent a pivotal field challenge, limiting the robustness of conventional feature-based localization and mapping approaches. This paper introduces VinePT-Map, a semantic mapping framework that leverages vine trunks and support poles as persistent structural landmarks to enable season-agnostic and resilient robot localization. The proposed method formulates the mapping problem as a factor graph, integrating GPS, IMU, and RGB-D observations through robust geometrical constraints that exploit vineyard structure. An efficient perception pipeline based on instance segmentation and tracking, combined with a clustering filter for outlier rejection and pose refinement, enables accurate landmark detection using low-cost sensors and onboard computation. To validate the pipeline, we present a multi-season dataset for trunk and pole segmentation and tracking. Extensive field experiments conducted across diverse seasons demonstrate the robustness and accuracy of the proposed approach, highlighting its suitability for long-term autonomous operation in agricultural environments.
Giorgio Audrito, Mauro Martini, Alessandro Navone +2
Department of Electronics and Telecommunications, Politecnico di Torino, 10129, Torino, Italy.
Accurate state estimation and multi-modal perception are prerequisites for autonomous legged robots in complex, large-scale environments. To date, no large-scale public legged-robot dataset captures the real-world conditions needed to develop and benchmark algorithms for legged-robot state estimation, perception, and navigation. To address this, we introduce the GrandTour dataset, a multi-modal legged-robotics dataset collected across challenging outdoor and indoor environments, featuring an ANYbotics ANYmal-D quadruped equipped with the Boxi multi-modal sensor payload. GrandTour spans a broad range of environments and operational scenarios across distinct test sites, ranging from alpine scenery and forests to demolished buildings and urban areas, and covers a wide variation in scale, complexity, illumination, and weather conditions. The dataset provides time-synchronized sensor data from spinning LiDARs, multiple RGB cameras with complementary characteristics, proprioceptive sensors, and stereo depth cameras. Moreover, it includes high-precision ground-truth trajectories from satellite-based RTK-GNSS and a Leica Geosystems total station. This dataset supports research in SLAM, high-precision state estimation, and multi-modal learning, enabling rigorous evaluation and development of new approaches to sensor fusion in legged robotic systems. With its extensive scope, GrandTour represents the largest open-access legged-robotics dataset to date. The dataset is available at https://grand-tour.leggedrobotics.com on HuggingFace (ROS-independent), and in ROS formats, along with tools and demo resources.
Turcan Tuna, Jonas Frey, Frank Fu +5
Robotic Systems Lab, ETH Zurich, Zurich 8092, Switzerland · 1ETH Zurich, Switzerland · 2Stanford University, USA +1