Fast Planar Grasp Pose Detection for Robot Based on Cascaded Deep Convolutional Neural Networks

A fast planar grasp pose detection method for robot based on cascaded convolutional neural networks is proposed to detect the pose for the unknown irregular objects with arbitrary poses. A cascaded two-stage convolution neural network model based on from coarse-to fine-scale position-attitude is established. The transfer-learning mechanism is used to train the model on small scale data sets. The grasp position candidate bounding-boxes are extracted and the coarse angle is estimated based on the R-FCN (region-based fully convolutional network) model. Angle-Net is proposed to solve the low accuracy detection problem of the previous methods, which can estimate the grasp angles with higher accuracy. Validations on the Cornell dataset and online grasp experiments on the real robot indicate that the proposed method can fast calculate the optimal grasp point and attitude for irregular objects with any shape and pose, the detection accuracy and speed are improved compared with the previous methods, the robustness and stability are strong, and it can be generalized to adapt to new object untrained.