A Cascaded Deep Learning Framework for Real-time and Robust Grasp Planning

Robotic grasping has been regarded as an essential way for robots to interact with the environment, which makes it a research hotspot. To our knowledge, current work on grasp planning could rarely take both robustness and time efficiency into account. To solve this problem, we propose a cascaded deep learning framework, which consists of a regression CNN and a refined CNN. Taking depth image as input, the first network regresses an optimal grasp region. The second network samples and sorts grasp candidates in the region and finally outputs a series of grasp poses with high quality. This architecture ensures real-time performance and acquisition of adequate candidates synchronously, which is especially appropriate for unstructured environment manipulation. We used Gazebo simulation and DART physics engines to build a grasp data generation environment. Comparing with manually annotating or robot collecting, our method makes data generation less expensive, more dynamic, and easier to generalize into the real world. The method proposed is evaluated in both simulation and physical experiments, verifying that our method is able to find multiple robuster grasp poses at a faster rate than the state of the art.