A 3-Stage Machine Learning-Based Novel Object Grasping Methodology

The automatic grasping of objects previously unseen by a robotic system is a difficult task-of which there is currently no robust solution. The research presented in this article improves upon previous works that employ depth data and learning techniques to generate and select from a pool of hypothesised grasps by focusing on the pruning and selection process. In this work, a vision-based, sampling methodology that generates candidate grasps through a convolutional neural network is proposed. Each candidate grasp is assessed using scores derived from the candidate itself and other related input modalities-such as the centre of gravity of the object. The final selection is determined by a learning algorithm. To overcome human bias, objective measures of grasp performance are established that comprehensively measure the error introduced by the grasp trial itself. The proposed metrics are empirically demonstrated to quantify grasp quality, offer useful criteria for network training and provide better descriptive power than traditional measures of grasp outcome. Experimentation showed that the proposed methodology can generate a meaningful, final grasp within 1.3 seconds. Trials quantitatively demonstrate a small-object-in-isolation performance of 99%. For unknown objects, this equates to a 10% improvement relative to other similar methodologies. Testing also showed that grasp performance was improved by 5% when implementing the proposed metrics-compared to the baseline.