Efficient Robotic Grasp Learning by Demonstration

In this paper, we propose a Learning from Demonstration approach for robotic grasping with compliant arms. The compliance in the robot arm for safety often causes a problem in grasping. In our approach, we construct a recurrent neural network, given the estimation of the target object position and random initial joint angles of the robot arm, that can produce the whole trajectories for grasping the target object. In order to generate smooth and stable trajectories and to decrease the number of human demonstrations, we propose a data augmentation method to increase the training data and utilize the trajectory planning technique using cubic splines for smooth and stable trajectories. Specifically, the two arms of the robot are trained respectively, and a support vector machine is used to decide which arm needs to be used for grasping the target object. The evaluation results show that our recurrent model not only has a good prediction for the final joint configurations, but also generates smooth and stable trajectory. Moreover, the model is robust to the changes in the initial joint state which means that even though the initial joint configuration is affected by disturbances, the model can still generate trajectories leading to the final joint configurations for grasping the object. Finally, we tested the proposed learning method on the Pepper robot which can successfully grasp randomly placed object on the workbench. Compared to traditional methods which need to avoid singular configurations as well as to secure accurate localization, our method turns out to be robust and efficient and can be applied to cluttered environment.