Grasp Synthesis Methods on Known Objects for Bin-Picking - A comparison on deep-learning based and analytical-model based 6 DoF grasp pose synthesisers

Abstract

Abstract Automated pick-and-place operations are foundational tasks in robotics, with wideranging applications in industrial automation, logistics, and service robotics. Central to these operations is the ability of a robotic system to reliably plan and execute grasps on a diverse set of objects. Grasp synthesis, which is the process of determining suitable contact points and hand configurations for successful object manipulation, remains a challenging problem due to the inherent uncertainties in perception, object variability, and physical interactions. To address the challenges of grasp synthesis, this thesis explores and evaluates two distinct approaches to grasp pose generation. The first approach leverages a database-driven method, storing precomputed grasp poses for known, proprietary objects. The second employs an end-to-end deep learning model capable of generalizing grasp predictions across a wide variety of novel objects. A complete robotic pipeline was developed to integrate these grasp synthesis methods into practical pick-and-place and bin-picking tasks. Using this pipeline, we conducted experimental evaluations on a physical robotic platform to compare the grasp success rates of both approaches in real-world scenarios. We conclude that the deep learning method, using Contact-GraspNet for generating grasps, appears more fitting for the applications Volvo desire in their production environment due to its flexible and scalable nature as well as achieving an overall 50% success rate compared to 39% for the database-driven method for single object pick-and-place. While the database-driven method could still work, it is not as scalable and is reliant on an object pose estimation system.