Domain Adaptation With Evolved Target Objects for AI Driven Grasping

• Published in: 2022 IEEE 27th International Conference on Emerging Technologies and Factory Automation (ETFA) pp. 1 - 8
• Main Authors: Pratheepkumar, Anish, Hofmann, Michael, Ikeda, Markus, Pichler, Andreas
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06-09-2022
• Subjects: Adaptation models; complex object evolution with diversity; Data models; domain adaptation in object grasping; Grasping; Prediction algorithms; Predictive models; robotic object grasping; Service robots; Training data
• DOI: 10.1109/ETFA52439.2022.9921470

The recent developments of AI and deep learning, together with increased data availability, have made noticeable progress in the field of robotic object grasping. However, the performance of state-of-the-art models in predicting reliable grasp for a given set of novel industrial objects, especially from a different domain with respect to the training data, still needs improvement. In this regard, we propose a novel approach of applying domain adaptation to the novel object synthesis by generating evolved objects. This proposed approach addresses the use case scenario of grasping a newly introduced set of industrial objects and an associated lack of training data. To realize the aforementioned domain adaptation, we apply genetic algorithm based on the method in Evolved Grasping Analysis Dataset (EGAD). Here we modify it to generate evolved objects from a given complex object rather than generating random objects, and the modified algorithm is referred to as EGAD-COMPLEX (EGAD-CMPLX). Our evaluation results show that for a given test set of novel target domain industrial objects, the grasp prediction model developed using the proposed evolved objects and the EGAD objects have superior performance. Specifically, the proposed model have, on average, 16 percent better grasp success rate than the baseline model.