This paper introduces a synchronous dual-arm manipulation with obstacle avoidance trajectory planning by an adult-size humanoid robot. In this regard, we propose a high precision 3D object coordinate tracking using LiDAR point cloud data and adopting Gaussian distribution into robot manipulation trajectory planning. We derived our 3D object detection into three methods included auto K-means clustering, deep learning object classification, and convex hull localization. Therefore, a lightweight 3D object classification based on a convolutional neural network (CNN) has been proposed that reached 91% accuracy with 0.34ms inference time on CPU. In empirical experiments, the Gaussian manipulation trajectory planning is applied adult-sized dual-arm robot, which shows efficient object placement with obstacle avoidance.
3D perception, deep learning, manipulation, humanoid robot