Grasping Unknown Objects Using Convolutional Neural Networks

Robotic grasping has been a prevailing problem ever since humans began creating robots to execute human-like tasks. The problems are usually due to the involvement of moving parts and sensors. Inaccuracy in sensor data usually leads to unexpected results. Researchers have used a variety of sensors for improving manipulation tasks in robots. We focus specifically on grasping unknown objects using mobile service robots. An approach using convolutional neural networks to generate grasp points in a scene using RGBD sensor data is proposed. Two convolutional neural networks that perform grasp detection in a top down scenario are evaluated, enhanced and compared in a more general scenario. Experiments are performed in a simulated environment as well as the real world. The results are used to understand how the difference in sensor data can affect grasping and enhancements are made to overcome these effects and to optimize the solution