Cooperative pepper picking: a Case Study on Yield Estimation in IndoorAgriculture

Marsela Polic, Antun Ivanovic, Matko Orsag

Keywords: deep learning, sim2real, heterogeneous robots

TL;DR: Heterogeneous robots deployed in a simulated greenhouse estimating sweet pepper yield using deep learning approach trained on a synthetic dataset.

Abstract: Our case study focuses on indoor robotic farming. Inspired by recent promising results in sim-to-real transfer we built a realistic simulation environment combining a ROS-compatible physics simulator (Gazebo) with a realistic rendering cycles engine from Blender. Without loss of generality, we focus on a sweet pepper harvesting task and showcase and analyze the technological pipeline necessary to conduct such a mission. The pipeline starts from aerial robotics control and trajectory planning, combined with deep learning-based pepper detection, a clustering approach for yield estimation, and mission planning for harvesting using a heterogeneous team of robots.