Artificial Neural Network with Bayes' Rule for Reasoning Task-Oriented Grasp

For the cognitive process of human grasping, the grasp position for a specific object is not only highly related to its inner structural morphology but also with its subsequent functional purpose of the intended manipulation task. In this paper, we propose a Bayesian reasoning improved neural network (BRNN) to identify the task-oriented grasp positions. In this model, the grasping task is input as a hyperparameter into BRNN to control the feature learning of object. A telescopic translator is learned to transform the task-oriented parameter distribution of the BRNN model to obtain the grasp positions of the current task based on the previous learned task. A task-switching mechanism is realized to switch the BRNN reasoning modes for different tasks. Experimental results show that when there is no task constraint, the model has the grasp detection accuracy of 95.2% on the Cornell Grasp Dataset, a little less than the state-of-the-art performance of 96.0%. However, with the task constraint, the model has the accuracy of 95.1% and 95.6% on the self-built BJUT-Grasp dataset for the reasoning of the non-function-use grasp tasks and function-use grasp tasks, respectively. For the visual 2D images, the BRNN model improves the reasoning ability for grasp detection and shows excellent ability in the task-oriented grasp position reasoning.