基于骨架型表示和域知识迁移的机器人杂乱环境抓取检测

As the progressive improvement of intelligence, the robot has been gradually applied to practical tasks. However, under the unstructured manipulation environments with complex, dynamic and uncertain conditions, the robotic grasp detection still suffers the following challenges: First, the graspable ability represented by current grasp representation is weak. Second, collecting and labeling the training samples are expensive and time-consuming. Third, the occlusion relationships between multiple objects are complicated, especially under cluttered scenarios. In order to solve these challenges, this project takes multi-object grasping under cluttered scenarios as research target and systematically studies the general grasp representation, grasp knowledge transfer as well as grasp affiliation modeling. The main research contents include: First, it builds the skeleton-based grasp representation to indicate more graspable regions and improve their evaluation accuracies. Second, the small-scale synthetic training samples are utilized to learn domain-invariant features, which decreases the model’s dependence to large amount of data and improves its adaptation to varying scenes. Third, the multi-granularity semantic segmentation is integrated to construct part-based grasp affiliations and boost the grasping logics under complex, occluded environment. Last, the researches will be applied to assist the practical experiment of robotic grasping. This project aims to form the algorithmic system about robotic grasp detection in cluttered scenes based on skeleton representation and cross-domain knowledge transfer, and further lays the theoretical foundation for the development of grasp perception system in real robotic manipulation tasks.

目前，随着智能化程度的不断提高，机器人正逐步走向实用化，然而在复杂、动态和不确定的非结构化操作环境下，机器人的抓取检测研究仍面临以下挑战：一是现有抓取表示方式的抓取表达能力差，二是带标签抓取样本的采集、标注成本高昂，三是杂乱环境下多物体间的遮挡关系复杂。本项目针对上述挑战，以杂乱环境下的多物体抓取为研究对象，对泛化抓取表示、抓取知识迁移和抓取隶属关系建模进行系统性科学研究，主要内容包括：建立覆盖完整抓取区域的骨架型抓取表示，提高稀疏可抓取区域的评价精度；利用小规模虚拟训练样本学习域不变抓取知识，降低抓取检测的数据依赖性，并提高其场景适应性；基于多粒度语义分割建立部件级抓取隶属关系，强化复杂遮挡环境下的抓取逻辑；并在真实机器人抓取系统下开展示范应用。本项目将形成一套基于骨架型抓取表示和虚拟域抓取知识迁移的机器人杂乱环境抓取检测方法体系，为机器人实际操作任务中抓取感知系统的研制奠定理论基础。