Grasp Pose Detection with Affordance-based Task Constraint Learning in Single-view Point Clouds
- PDF / 7,419,986 Bytes
- 19 Pages / 595.224 x 790.955 pts Page_size
- 22 Downloads / 198 Views
Grasp Pose Detection with Affordance-based Task Constraint Learning in Single-view Point Clouds Kun Qian1,2 · Xingshuo Jing1,2 · Yanhui Duan1,2 · Bo Zhou1,2 · Fang Fang1,2 · Jing Xia1,2 · Xudong Ma1,2 Received: 19 October 2019 / Accepted: 13 April 2020 © Springer Nature B.V. 2020
Abstract Learning to grasp novel objects is a challenging issue for service robots, especially when the robot is performing goaloriented manipulation or interaction tasks whilst only single-view RGB-D sensor data is available. While some visual approaches focus on grasping that satisfy force-closure standards only, we further link affordances-based task constraints to the grasp pose on object parts, so that both force-closure standard and task constraints can be ensured. In this paper, a new single-view approach is proposed for task-constrained grasp pose detection. We propose to learn a pixel-level affordance detector based on a convolutional neural network. The affordance detector provides a fine grained understanding of the task constraints on objects, which are formulated as a pre-segmentation stage in the grasp pose detection framework. The accuracy and robustness of grasp pose detection are improved by a novel method for calculating local reference frame as well as a position-sensitive fully convolutional neural network for grasp stability classification. Experiments on benchmark datasets have shown that our method outperforms the state-of-the-art methods. We have also validated our method in real-world and task-specific grasping scenes, in which higher success rate for task-oriented grasping is achieved. Keywords Robot grasp · Grasp pose detection · Object affordance · Convolutional neural networks · Constraints learning
1 Introduction Knowing how to grasp is generally more challenging than what to grasp for a robot. In the real world, planning a grasp in a single-view point cloud is challenging [18, 36], because a grasping model that can work with raw sensor input is needed to account for the uncertainty brought about by inaccurate and incomplete RGB-D sensing. In situations that the CAD model of the objects are unknown, Grasp Pose Detection(GPD) methods attempt to detect 3DoF or 6DoF grasp poses directly from RGB-D sensor data without the estimation of the object poses. Therefore, the GPD methods have yielded promising results across a range of objects and their models generalize well to novel objects.
Kun Qian
[email protected] 1
School of Automation, Southeast University, Nanjing 210096, China
2
Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education, Nanjing 210096, China
While the Grasp Pose Detection (GPD) methods [8, 38] generate stable grasps on any objects, in this paper we focus on task-specific (i.e., goal-oriented) grasping in object manipulation as well as human-robot interaction scenarios, where robots are supposed to grasp objects or tools first before executing a task or interacting with humans. To reason about task requirements and to satisfy task-specific constrai
Data Loading...
