Robotic grasp manipulation is the ultimate stage of effective human robot interactions prior to scene understanding, localization, receptacle, and object detection. An intelligent robot endeavors to replicate the intricate dexterity exhibited by human hands, necessitating an intelligent grasping ability for executing complex tasks, which is much superior compared to earlier practiced planning based strategies, especially when the robot needs to work in an unstructured environment. However, teaching robots to grasp objects effectively poses significant challenges. Unlike humans, who refine this skill through years of trial and error during their child hoods, robots lack the luxury of a prolonged learning period. To expedite their learning process, in this chapter, we are proposing to utilize the deep learning and machine learning techniques, exploiting both generative AI and discriminative AI for imparting intelligence to robot grasping. This chapter presents our recent research and its application in this exciting domain of vision based robotics. Although we have presented our works on tabletop environments, the similar strategies can be scaled up for 6-D pose as well. Given the data-intensive nature of deep learning and the scarcity of labeled data, we advocate for a hybrid approach that integrates discriminative and generative models employing the vector quantized variational autoencoder (VQ-VAE). We present the development of two cascaded: Representation based Generative Grasp Convolution Neural Network (RGGCNN) and Representation based Generative Grasp Convolution Neural Network 2 (RGGCNN2), and two advanced models: the Generative Inception Neural Network (GI-NNet) and the Representation based GI-NNet (RGI-NNet). GI-NNet excels in generating antipodal robotic grasps, achieving an impressive 98.87% accuracy on RGB-Depth images sourced from the Cornell Grasping Dataset (CGD). RGI-NNet, incorporating VQ-VAE, demonstrates robust learning capabilities even with limited labeled data, achieving grasp pose accuracies ranging from 92.13% to 97.75%. To validate the efficacy of our models, we conduct experiments utilizing Anukul (Baxter) Cobot for real-time tabletop grasp executions. As a part of our future endeavors, we aspire to integrate our grasping strategies with the open vocabulary mobile manipulation framework, thereby facilitating natural language interactions between humans and robots, which could significantly enhance the usability and versatility of robotic systems in various real-world scenarios.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    Estimating Grasp Pose Using Deep Learning Architectures for Intelligent Robotic Manipulation


    Contributors:

    Published in:

    Publication date :

    2024-07-24


    Size :

    30 pages




    Type of media :

    Article/Chapter (Book)


    Type of material :

    Electronic Resource


    Language :

    English