GSoC’21 RoboComp project: Deep Reinforcement Learning for precise robot grasping and object manipulation

19th June, 2021

About me

I am an undergrad student, pursuing a Bachelor’s degree in Computer Science Engineering with specialization in Artificial Intelligence and Machine Learning along with a minor degree in Robotics from SRM Institute of Science and Technology, Chennai, India. I have worked on multiple robotics projects and have developed and deployed two 6 DOF and one 3 DOF robotic arm as a part of them.

About the Project

The project aims to use the approach of deep reinforcement learning to create a pipeline for robot control with object grasping and manipulation. Grasping an object is the most fundamental objective of robotic end effectors. Furthermore, every object manipulation task begins with object grasping. The methods deployed currently for this depend on accurate mathematical models of the robot arm. However, the approach this project aims is to teach the arm to accomplish the task on its own similar to how a baby learns to walk or ride a bicycle called reinforcement learning. This can be achieved with either an end-to-end approach (using rgb signals from a camera) or using currently implemented object perception modules (DNN estimated poses). Upon successful development of the workflow, it needs to be tested on a Kinova Gen3 arm.

Pawan Wadhwani