Versatile grasping for shelf placement of FMCG items

• Main Author: Das Bisakha
• Other Authors: Sourav Sen Gupta
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/153312
• Institution: Nanyang Technological University

Versatile grasping is one of the most basic forms of robotic manipulation. Versatile grasping's purpose is to gain great autonomy in dexterous manipulation tasks in an unstructured environment. An example of such an unstructured environment with scope of versatile pick and placement options is the task of shelf-placement, requiring high and low levels of perceptual reasoning. In this paper, we chose to extend the novel model structure of Transporter Networks beyond tabletop actions performed on extruded 2-dimensional solid objects utilizing 3 Degrees of Freedom (DoF). These extensions are based on the development of an agent capable of performing shelf- placement with 6-DoF movements on two distinct axis planes. This agent has been trained on a dataset of weighted Fast-Moving Consumer Goods (FMCG) objects, both un-textured and textured. Since the training was based on imitation learning, an expert agent was developed and implemented as well. The results obtained from training the 6-DoF agent on demonstrations provided by the expert agent confirm its successful extension to 6-DoF on two planes of axis with a 70% accuracy on FMCG products. The results further indicate the success of the agent on industry benchmarked untextured and textured Yale-CMU-Berkeley (YCB) objects. In addition to extending and contributing to existing research, this work also paves the way for future research with a real UR5e robot.