Novel Objects Detection for Robotics Grasp Planning

Abstract

Object detection and segmentation have seen significant advances in robotics grasp planning in the last few years. But many algorithms and methods are still in low accuracy and unsuitable for real-time performance. This paper presents a data-driven approach to perform objects and grasp point detection from camera images for novel objects in cluttered environments. To achieve this, it introduces a model with two parallel loops of image encoders and decoders, to perform object detection, segmentation, and object grasping points detection between raw mages and robotics grasping process simultaneously. The experimental results demonstrate that our model and algorithm achieve efficiently grasping performance for various types of novel objects with high success rates in a cluttered environment. We improved the detection accuracy compared to the basic ResNet and DeepLab model by more than 11 % rate and have better performance in grasping detection than other comparable methods.