Reinforcement Learning for Robotic Grasping and Manipulation: A Review

摘要

A century of robots is the 21st century. The robots have long been able to cross the divide between the virtual universe and the real world. Robotics, as the most successful contender in the upcoming great technological revolution, will play an ever more important role in society because of the impact it has in every field of life, including medicine, healthcare, architecture, manufacturing and food supplies, logistics and transport. This document introduces a modern approach to the grasp of robots, which draws grasp techniques from the human demonstration and combines these strategies into a grasp-planning framework, in order to produce a viable insight into the objective geometry and manipulation tasks of the object. Our study findings show that grasping strategies of the form of grasp and thumbs positioning are not only necessary for human grasp but also significant restrictions on posture and wrist posture which greatly reduce both the robot hand's workplace and the search space for grasp planning. In the simulation and with a true robotic system this method has been extensively tested for several everyday living representative objects. In the experiment with varying degrees of perceiving in certainties, we have demonstrated the power of our method.