In this paper, we present a learning platform for robotic grasping in real world, in which actor-critic deep reinforcement learning is employed to directly learn the grasping skill from raw image pixels and rarely observed rewards. This is a challenging task because existing algorithms based on deep reinforcement learning require an extensive number of training data or massive computational cost so that they cannot be affordable in real world settings. To address this problems, the proposed learning platform basically consists of two training phases; a learning phase in simulator and subsequent learning in real world. Here, main processing blocks in the platform are extraction of latent vector based on state representation learning and disentanglement of a raw image, generation of adapted synthetic image using generative adversarial networks, and object detection and arm segmentation for the disentanglement. We demonstrate the effectiveness of this approach in a real environment.

The optimal grasping point of the object varies depending on the shape of the object, such as the weight, the material, the grasping contact with the robot hand, and the grasping force. In order to derive the optimal grasping points for each object by a three fingered robot hand, optimal point and posture have been derived based on the geometry of the object and the hand using the artificial neural network. The optimal grasping cost function has been derived by constructing the cost function based on the probability density function of the normal distribution. Considering the characteristics of the object and the robot hand, the optimum height and width have been set to grasp the object by the robot hand. The resultant force between the contact area of the robot finger and the object has been estimated from the grasping force of the robot finger and the gravitational force of the object. In addition to these, the geometrical and gravitational center points of the object have been considered in obtaining the optimum grasping position of the robot finger and the object using the artificial neural network. To show the effectiveness of the proposed algorithm, the friction cone for the stable grasping operation has been modeled through the grasping experiments.

This paper proposes a method to simultaneously estimate two degrees of freedom in wrist forces (extension - flexion, adduction - abduction) and one degree of freedom in grasping forces using Electromyography (EMG) signals of the forearms. To correlate the EMG signals with the forces, we applied a multi - layer perceptron(MLP), which is a machine learning method, and used the characteristics of the muscles constituting the forearm to generate learning data. Through the experiments, the similarity between the MLP target value and the estimated value was investigated by applying the coefficient of determination (R2) and root mean square error (RMSE) to evaluate the performance of the proposed method. As a result, the R2values with respect to the wrist flexionextension, adduction - abduction and grasping forces were 0.79, 0.73 and 0.78 and RMSE were 0.12, 0.17, 0.13 respectively.