Simulation-Based Synthetic Data Generation for Automated Training of Bin-Picking Segmentation Models

Abstract

Object segmentation is critical for addressing binpicking challenges in robotics, where scenes often involve overlapping, partially occluded objects. Accurate segmentation facilitates precise object localisation and identification. In this work, we introduce a Unity-based tool with domain randomisation functionalities for generating synthetic cluttered bin-picking scenes with automatic pixel-level annotations. We validated the usefulness of the system in two real use cases: the manipulation of vegetables and mechanical parts. By allowing models to be trained entirely on synthetic data, this solution eliminates the need for costly and time-consuming real-world data collection and annotation.