Memory Efficient Grasping Point Detection of Nontrivial Objects

Robotic manipulation with a nontrivial object providing various types of grasping points is an industrial interest. Here, efficient method simultaneous detection the proposed. Specifically, two different 3 degree-of-freedom end effectors are considered for grasping. The utilizes RGB data-driven perception system based on specifically designed fully convolutional neural network called attention squeeze parallel U-Net (ASP U-Net). ASP detects single image. This image transformed into schematic grayscale frame, where positions and poses coded gradient geometric shapes. In order to approve architecture, its performance was compared nine competitive architectures using metrics generalized intersection over union mean absolute error. results indicate outstanding accuracy response time. also computationally enough. With more than acceptable memory size (77 MB), architecture can be implemented custom single-board computers. capabilities were tested evaluated NVIDIA Jetson NANO platform.