This paper presents a new approach to robotic manipulation, focusing on bin-picking. It leverages category-agnostic instance segmentation and simulation-based training for real-world application. The method can handle objects regardless of their class, accommodating even transparent and semi-transparent items. The strategy overcomes challenges from noisy depth sensors, enhancing learning reliability. It includes domain randomization for successful transfer, a dataset for warehouse applications, and an efficient bin-picking framework. The model achieves state-of-the-art performance over the WISDOM public benchmark and a custom dataset, with 98% accuracy for opaque objects and 97% for non-opaque objects.
Publication date: 29 Dec 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2312.16741
