Machine Learning Reveals Male and Female Yerba Mate Trees Read Light in

Researchers have tested a robot that picks oranges in a California orchard. They found that a simple V-shaped guide and carefully planned turns of the robot arm can replace a more complicated motor-powered system for lining up the cutting blades. The robot can still cut orange stems short and neatly, which is important for oranges sold fresh in stores rather than used to make juice. In tests with 100 navel oranges, the robot picked every orange without damaging it. The average length of stem left on the oranges was 3.10 ± 3.34 millimeters, and the middle value was 1.81 millimeters. About 92 percent of the oranges had less than 10 millimeters of stem left, meeting a commonly used quality standard. Another 52 percent met a stricter standard of less than 2 millimeters. The researchers, from California State University, Fresno, say this method could help robots harvest oranges that are good enough to sell fresh.

Picking oranges for grocery stores is difficult because the remaining stem must be very short. If a stem is too long, it can poke holes in other oranges during washing, packing, or shipping. But cutting too close to the orange's skin can damage the fruit. Damage where the stem connects to the orange may mean the fruit must be used for juice instead of being sold fresh. This is one reason why commercial orange-picking robots have mostly been used to harvest fruit for juice. To become useful for fresh oranges, robots must be able to control how much stem they leave behind.

However, most existing orange-picking robots do not even measure the length of stem left behind. When the Fresno researchers looked at other robotic systems that cut citrus stems, they found that most studies reported only how often the robots successfully picked fruit and how long picking took. Few systems carefully lined up the stem with the cutting blades before making the cut. Comparing picking times between studies is also difficult. Some studies count only the time used by the picking tool, while others include finding the fruit, moving toward it, carrying it away, and returning the robot arm. The Fresno team's earlier robot used a flexible gripper and motor-controlled cutting shears mounted on a platform that could move in two directions. This platform moved the blades into position around the stem. In separate tests where a person positioned the tool at each orange, the system left an average stem length of 0.81 ± 0.46 millimeters. It was very accurate, but the moving platform needed extra motors, sensors, weight, and complicated controls.

The new robot works in the opposite way. Instead of moving the cutting blades toward the stem, it moves the stem toward the blades. Its picking tool has three flexible fingers made from a soft plastic called thermoplastic polyurethane. A small motor moves a four-bar mechanism that opens and closes a pair of commercial citrus cutting shears. The blades curve to match the orange's shape, allowing them to cut the stem close to the skin. A V-shaped guide sits between the fingers and can handle stems pointing as much as 30 degrees away from the correct position in any direction. After grabbing an orange, the robot turns its tool 10 degrees clockwise and then 10 degrees counterclockwise. Next, it turns the arm's fourth joint downward by 5 to 10 degrees. These movements help the stem slide into the V-shaped guide and settle near the bottom of the cutting blades. The fingers then close completely, and the shears cut the stem.

Using the robot's movements instead of extra moving parts made the picking tool much simpler. The new tool weighs only 0.5 kilograms, compared with 1.5 kilograms for the earlier design. That is a 67 percent reduction in weight. The number of motors and other powered moving devices also dropped from four to two. The researchers removed two stepper motors, their controllers, two switches that detect movement limits, and a pressure sensor. They also removed about 7.2 seconds of powered platform movement from each picking cycle. That movement had taken up 67 percent of the earlier tool's operating cycle. The new tool's average operating time fell from 10.79 seconds to about 9.1 seconds. During orchard tests, the robot successfully picked all 100 oranges, giving it a 100 percent success rate for removing fruit. None of the oranges had holes, bruises, or scraped skin.

The complete harvesting system includes more than the picking tool. A mobile vehicle made by Bonsai Robotics carries a Fairino FR3 robot arm that can move in six different ways, along with an Intel RealSense D457 camera that captures both color pictures and depth information. The researchers used an artificial intelligence system called YOLOv8 to find oranges and their stems in camera images and identify their outlines. They trained the system using 1,159 labeled images taken at the Fresno Campus Farm orchard. They expanded these into 2,790 training images by creating additional versions. The system combines color images with depth measurements to work out where each orange is in three-dimensional space. In a separate test set, the AI reached a mean average precision of 95.5 percent for oranges and 53.3 percent for their thin stems, using an intersection-over-union threshold of 0.50. This measure checks how closely the AI's predicted outlines match the correct outlines. The researchers explain that small objects like stems are especially affected by small mistakes along their edges. The V-shaped guide handles the final positioning of the stem, so the camera system does not have to be perfectly accurate. In orchard tests involving 82 oranges, the vision system detected 95.1 percent of the oranges and 87.8 percent of their stems.

The camera system was accurate enough to help the robot find, grab, and center the oranges. Across 47 samples, its average error when finding an orange's position was 33.27 ± 15.27 millimeters. The biggest part of this error was sideways, averaging 22.53 ± 16.37 millimeters. The researchers believe this happened because leaves, branches, or other objects sometimes blocked part of an orange, making its estimated center appear to be in the wrong place. The system also estimated each orange's diameter by treating it as a sphere. These size estimates were off by an average of only 4.07 ± 4.40 millimeters, or 3.81 percent. The average errors when estimating the direction of the fruit and stem were 3.1 ± 4.22 degrees and 5.9 ± 7.89 degrees, respectively. Every error in estimating stem direction stayed within the V-shaped guide's 30-degree limit. The average orange picked was 73.52 millimeters across, larger than the average error in finding its position. Because the gripper's flexible fingers could also adjust to small positioning mistakes, the robot did not fail to approach fruit because of depth-measurement errors in sunny or partly cloudy daylight.

The robot took an average of 55.1 seconds to complete the entire picking process for one orange. The researchers deliberately ran the arm at only 20 percent of its maximum rated speed so a human operator could step in when needed. Looking at each part of the process showed that carrying an orange to the collection bin took 15.52 seconds, while moving the arm back to its starting position took 15.37 seconds. Together, robot arm movements made up about 66 percent of the total picking time. The picking tool itself worked much faster. Grabbing the orange and moving its stem into position took 7.59 seconds, and cutting took only 1.53 seconds. Finding the fruit took just 0.40 seconds. The arm followed a set path through predetermined positions and did not plan new routes to avoid obstacles. During 10 of the 100 picking cycles, problems with movement interrupted the process. These included touching branches, reaching arm positions that were difficult to control, and controller errors. Each interruption required help from the operator, although the robot still successfully removed the orange every time.

The researchers also looked closely at why the new robot left an average stem length of 3.10 millimeters, compared with just 0.81 millimeters for the earlier design. All 48 oranges with more than 2 millimeters of stem remaining had at least one problem that made cutting harder. These included uneven or bent stems, stems pointing at extreme angles, stems with branches that got in the way, or nearby branches and oranges. The other 52 oranges, which had none of these problems, had an average remaining stem length of 1.01 ± 0.56 millimeters. This was close to the earlier robot's average of 0.81 millimeters. The difference of about 0.2 millimeters was statistically significant, meaning it was unlikely to be due to chance alone. It showed the small loss of cutting accuracy that came from removing two motors used for positioning when conditions were otherwise comparable. After accounting for how many oranges belonged to each group, the researchers found that the 48 harder-to-pick oranges explained 2.18 millimeters, or 95 percent, of the 2.29-millimeter difference between the average stem lengths in the two studies. Importantly, mistakes in the camera system were not responsible. There was no connection between errors in estimating stem direction and the length of stem left behind. When the guide did not center a stem perfectly, the robot simply left a longer piece of stem rather than failing to pick the orange. The researchers warn that this comparison used results from separate earlier and later tests. A controlled experiment using both picking tools on the same oranges is still needed to find out exactly how much the different positioning methods affect cutting accuracy.

The researchers also point out the limits of their study. All tests used navel oranges from one orchard during a single harvest season. The oranges were between 80 and 120 centimeters above the ground, and the tests happened in daylight. A human operator watched the robot throughout the tests, drove its vehicle, chose the arm's starting positions, and started each picking cycle. The team has not tested how well the system works with mandarins, lemons, trees with different branch arrangements, or low-light conditions. The researchers also did not study the costs and financial benefits of using the robot. In its current partly automatic form, which still needs human supervision, the system does not save labor compared with picking oranges by hand. However, the team has several ideas for improving it. The camera system already estimates which direction a stem points, so the robot could use that information to turn the picking tool toward the stem before approaching. It could also adjust how far it rotates the fruit based on the stem's estimated angle. Using images from several viewpoints could help it see stems hidden behind other objects. Better movement planning could help the arm avoid branches and other obstacles. The robot could also use electrical current readings from its existing motors to check automatically whether it had grabbed and cut an orange successfully. The researchers already recorded these signals during testing. Early tests at higher robot speeds reduced the average picking time to 38.6 ± 10.8 seconds per orange, but the arm stopped for safety more often when it touched branches. Overall, the results show that a simple V-shaped guide can help a robot control how much orange stem remains after cutting, without needing complicated extra machinery. Better use of stem direction information and smarter arm movements could make the robot faster and more reliable.

Subject of Research: A robotic citrus-picking tool that uses a simple guide to center stems and control how much stem remains after cutting.

Article Title: Passive stem-centering end-effector for controlled residual stem length: Integration and orchard evaluation on a robotic citrus harvesting platform

Article References: Chinnock, B., Catalan, A., Yang, D., Oliva, N., Ellis, A., Nguyen, T., & Bani Milhim, A. (2026). Passive stem-centering end-effector for controlled residual stem length: Integration and orchard evaluation on a robotic citrus harvesting platform. Smart Agricultural Technology, 15, Article 102646. https://doi.org/10.1016/j.atech.2026.102646

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102646

Keywords: robotic harvesting, citrus, stem-centering, agricultural robotics, end-effector, navel oranges, machine vision, YOLOv8, fresh-market fruit, residual stem length, orchard automation, compliant gripper