Predicting the robot's grip capacity on different objects using multi-object grasping

Abstract

This study explores the novel concept of Multi-Object Grasping (MOG) and develops an architecture based on autoencoders and transformers for accurate object prediction in MOG scenarios. The approach employs different deep learning methods and diverse training approaches using the ping pong ball dataset. The parameters obtained from this training enhance the model's performance on the actual system dataset, serving as the final test and validation of the model's reliability in real-world situations. Comparing the model's performance on both datasets facilitates validation and refinement, affirming its effectiveness in practical robotic applications. The study highlights that training various dataset features significantly improves prediction accuracy compared to the Naïve model using dense neural networks. Using five-time steps notably enhances prediction accuracy, especially with the GRU model in time-series data architecture, achieving a peak accuracy of 96%. While MOG has been extensively studied, this study introduces a novel architecture distinct from traditional visual methods. A framework is established that utilizes autoencoder and transformer technologies for managing tactile sensors, hand pose joint angles and force measurements. This approach demonstrates the potential for accurately predicting multiple objects in MOG scenarios.