HortiBot: An Adaptive Multi-Arm System for Robotic Horticulture of Sweet Peppers
This presentation explores HortiBot, a three-arm robotic system designed to automate labor-intensive horticultural tasks like selective harvesting of sweet peppers. The talk examines how the system integrates active perception with dual-arm manipulation to navigate dense greenhouse environments, achieve precise peduncle cutting, and adapt in real-time to variable plant structures. We'll see how this approach addresses workforce shortages while demonstrating a general-purpose platform adaptable to multiple horticultural operations beyond harvesting alone.Script
In greenhouses across the world, skilled workers carefully prune stems and selectively harvest peppers by hand, a task so delicate and variable that automation has remained frustratingly out of reach. Now a three-arm robot called HortiBot is changing that equation.
Previous robotic systems struggled with a fundamental tension: greenhouses are structured enough to seem automatable, yet plant density and biological variation make every manipulation unique. The authors recognized that solving this requires more than clever grippers or better vision.
Their answer is a system where perception and manipulation work as partners, not separate stages.
HortiBot uses three arms with distinct roles. While two arms manipulate the plant, a third arm continuously repositions stereo cameras to refine the robot's understanding of where the peduncle, the slender stalk connecting fruit to stem, actually is. This active perception loop turns what was a blind, pre-planned motion into an adaptive dance.
Earlier robots like SWEEPER and Harvey captured the scene once, then executed blindly. HortiBot instead treats perception as an ongoing conversation: as one arm moves a branch aside, the perception arm shifts to see the newly exposed peduncle, updating the cutting trajectory on the fly. The robot literally looks where it's reaching.
The devil is in the peduncle, a structure often just millimeters wide and partially occluded by leaves. By cropping the camera view tightly around each detected pepper, the system dramatically improves its ability to locate this critical cutting point. Force sensors then confirm contact, preventing the gripper from crushing what it's trying to harvest.
In trials with actual pepper plants, HortiBot delivered higher success rates and faster cycle times than prior systems. The key wasn't just better hardware, but the architectural decision to make perception and action interdependent. When the robot can see what it's doing while it's doing it, precision compounds.
The authors are transparent about remaining obstacles. When peduncles hide deep within foliage or lighting creates harsh shadows, even active perception struggles. And coordinating three arms in real-time demands robust software architecture that can handle sensor noise and timing jitter without cascading failures.
HortiBot isn't just a pepper harvester. The authors framed it as a flexible platform where task-specific tools and perception strategies can be swapped in. This positions the system as a step toward general-purpose horticultural automation, where one robot adapts to seasonal tasks rather than requiring separate machines for each job. That adaptability could finally make robotic horticulture economically viable at scale.
What started as a harvesting challenge became a lesson in robot architecture: when you give a robot the ability to look while it acts, you don't just improve accuracy, you unlock entirely new categories of tasks. Visit EmergentMind.com to explore more research like this and create your own video presentations.
