Quadruped Robotics
for Agricultural Sensing

This study examines how well fixed greenhouse sensors represent the microclimate across crop rows and canopy heights. A quadruped robot measured temperature, relative humidity and CO₂ to map spatial variation and compare mobile observations with fixed-sensor references.

Materials & Methods

The experiment took place in a cherry tomato glasshouse in Miryang, Republic of Korea. The robot surveyed eight 54 m pipe-rail lanes: L1–L4 beside dense canopies and L5–L8 beside pruned canopies. The two canopy conditions provided contrasting environments for interpreting the mapped measurements.

A Unitree Go2 EDU carried identical environmental sensing modules at the fruit zone (1.15 m) and upper canopy (1.65 m). Each module combined a BME280 temperature/humidity sensor and a CozIR-LP2 CO₂ sensor. An I²C multiplexer, Arduino Pro Mini and USB-to-TTL interface connected the two modules to the acquisition computer. Environmental readings were recorded at 2 Hz per height.

LiDAR-based rail detection and PID steering guided the forward measurement runs at 0.5 m/s. ROS 2 recorded the environmental and robot data against a common clock. Four sessions—morning, midday, afternoon and evening—were conducted once each on separate days between 9 and 12 June 2026.

Sensor cross-calibration aligned the two modules, and fixed-sensor time series were used to correct temporal variation during sequential lane surveys. Observations were averaged into 160 measured cells per map (eight lanes × 20 along-lane intervals), yielding 24 maps across three variables, two heights and four sessions. Along-lane coordinates were estimated from elapsed traversal time and lane length; no spatial interpolation was used to fill the maps.

Results

Spatial variation was consistently structured. All 24 maps showed positive spatial autocorrelation (Moran’s I 0.40–0.85; p < 0.001). Upper-canopy relative humidity exceeded fruit-zone values by 2.6–4.4 percentage points, and CO₂ was 12–69 ppm higher, in every session. At the upper canopy, the pruned section was warmer and drier than the dense section; canopy condition and location were confounded, so the difference cannot be attributed to pruning alone.

The spatial overlay connects environmental readings with crop-row geometry. Figure 6 locates the two-height measurements within the greenhouse structure and makes differences between lanes and canopy levels easier to interpret. Measurement positions are relative lane coordinates reconstructed from traversal time, rather than independently validated absolute positions.

Fixed readings did not fully represent the mobile spatial distributions. Humidity references were 2.4–7.3 percentage points below mobile means; CO₂ references were 48–116 ppm lower except in the evening. Depending on the session, 19–53% of cells differed from the fixed temperature reference by more than 0.5 °C, and 18–76% differed from the humidity reference by more than 5 percentage points.

The results support periodic mobile surveys to assess fixed-sensor representativeness and identify locations needing closer monitoring. They do not establish absolute sensor accuracy, energy savings or yield effects. Because the four sessions occurred on different days, session differences reflect both time of day and day-specific conditions.

My contribution

As a co-first author, I contributed to the methodology, experimental investigation and manuscript preparation.

My M.S. thesis investigates how field-scale UAV imagery and close-range quadruped RGB-D observations can describe complementary soybean traits before harvest. A drone maps the upper canopy across the field, while the robot observes plant bases and pods below the canopy. The study links these observations in common 1 m row segments and tests the robot estimates against manual measurements.

Field platform and data collection

Ground platform: Unitree Go2 · RealSense D405 · MTi IMU · ZED-F9P RTK-GNSS · Jetson AGX Orin.
Aerial platform: DJI Mavic 3M with RGB and multispectral cameras. The soybean field is at the National Institute of Crop Science in Miryang, Republic of Korea.

UAV mapping with Agisoft Metashape

UAV imagery from 2 and 12 October 2026 was processed in Agisoft Metashape to reconstruct the field surface and produce RGB orthomosaics and digital surface models (DSMs). The outputs were resampled to a common local grid and oriented along the crop ridges for comparison with the robot surveys.

How the CHM was calculated. Canopy height is the DSM surface elevation minus the estimated local ground elevation. The initial Metashape ground classification included canopy in closed areas, so the ground model was rebuilt from low points in the 12 October DSM: the lowest 2% within 2 m × 2.8 m blocks, followed by removal of elevated blocks, interpolation and smoothing. These heights still require field validation and can reflect both plant height and leaf loss.

Selecting contrasting zones for robot surveys

These maps guide the next ground campaign. Six 10 m candidate zones, including the previously surveyed r03 area, have been identified across the field's canopy-height and maturity gradient. Selection uses the 12 October 90th-percentile canopy height (CH_p90) and the 2 October RGB green fraction, choosing non-overlapping zones with relatively uniform conditions while covering the range from early leaf loss to retained green canopy. The 2 October CHM is excluded because its surface is smoothed and its vertical offset is inconsistent.

Connecting field maps to individual plants

Latest robot analysis: plant tracking and pod height

The 12 October r03 survey covered about 9.4 m of mature soybean row. SAM3 plant tracking and PodNet pod detection were combined with measured D405 depth and registered camera poses to reconstruct visible plants and pods. Unlike the earlier presentation's DA3 examples, this result uses the depth sensor directly.

Low stems were used to correct plant bases and merge split tracks. Visible pods outside the plant masks were assigned to nearby bases in 3D. This increased the assigned share of valid pods from 80% to 95%.