Deep Robotic Prediction with hierarchical RGB-D Fusion
share
Robotic is a fundamental operation in robotic control task goals. We consider problem by multimodal hierarchical fusion to generate policy with partial observation. The most of current methods for robotic focus on RGBD policy in the table surface scene or 3D point cloud analysis inference in 3D space. Comparing to these methods, we propose a novel multimodal hierarchical deep method that fuses RGB and depth data to realize robotic humanoid in 3D space with only partial observation. Under supervised learning, we develop a general label algorithm to label ground-truth in common RGBD dataset. A real-time hierarchical encoder-decoder neural network is designed to generate policy. We evaluate the effectiveness and performance of our method on a physical robot setup. The results of experiment validate our method.
READ FULL TEXT