AGRIST
Picking peppers, one fruit at a time
Perception and end-effector work for an autonomous bell pepper harvesting robot built by AGRIST, a deep-tech agriculture startup in Miyazaki, Japan.
Overview
AGRIST is a deep-tech company based in Shintomi, Miyazaki, building affordable harvesting robots for Japanese greenhouse agriculture. Their flagship product is a rail-guided robot that autonomously locates and picks bell peppers — a crop that ripens unevenly and currently demands a brutal amount of stoop labor from a shrinking farming workforce.
I worked with the AGRIST team on the perception and end-effector side of the system: how the robot decides which pepper to take, and how the gripper closes around fruit it can only partially see through a canopy of leaves and stems.
The problem with picking peppers
Bell peppers are deceptively hard to harvest:
- Fruit hangs at irregular angles, often occluded by leaves or neighbouring peppers
- A pepper is only ripe in a narrow window — pick it too early and the plant’s yield drops, too late and it rots on the vine
- The stem must be cut cleanly without bruising the fruit or damaging the plant
- Greenhouses are hot, humid, dimly lit, and full of pollen — punishing conditions for both humans and cameras
A general-purpose manipulator with a generic gripper does poorly here. The system has to be tuned to the geometry of the crop and the layout of the greenhouse.
The robot
AGRIST’s robot runs on overhead pipe rails (Φ31.8–48.6mm, the same rails greenhouse heaters slide on), which sidesteps the navigation problem entirely. A single onboard camera scans the canopy as the robot crawls along the row at 1–10 cm/s. When it spots a ripe pepper, the arm extends, closes around the fruit, and severs the stem — about one pepper per minute.
What I worked on
The interesting research question is the loop between seeing and grabbing. A bell pepper is roughly a deformable cup with a thin woody stem on top; the gripper has to find a pose that captures the fruit without snagging on adjacent peppers, then locate the stem and cut. Two problems sit underneath this:
- Perception under occlusion. Training the detector on idealised, well-lit fruit gives misleading accuracy numbers. Real harvests happen against a dense green background with peppers half-hidden behind leaves and other peppers. The detector has to output not just “pepper here” but a usable estimate of orientation and stem location from a partial view.
- Designing the gripper around the perception. A more capable gripper relaxes the demands on vision; a smarter vision system relaxes the demands on the gripper. The two have to be co-designed rather than handed off across a clean interface.
The work was less about building a single component and more about iterating across that boundary — prototyping grippers, collecting greenhouse data, retraining, and watching what actually broke when the robot ran in a real farm.
Context
AGRIST has been recognised by the UN Development Programme, partners with Microsoft on the image-analysis side, and is one of the few startups attacking the labour crisis in Japanese protected agriculture from the robotics end rather than the software-only end. Working with them was a chance to put research-quality perception against a problem where every failure costs a real fruit on a real farm, not a number on a test set.
Learn more about AGRIST at agrist.com.
More Images
