In robot programming by demonstration (PbD) of small parts assembly tasks, the accuracy of parts poses estimated by vision-based techniques in demonstration stage is far from enough to ensure a successful execution. This paper aims to develop an inference method to improve the accuracy of poses and assembly relations between parts by integrating visual observation with computer-aided design (CAD) model.,In this paper, the authors propose a spatial information inference method called probabilistic assembly graph with optional CAD model, shorted as PAGC*, to achieve this task. Then an assembly relation extraction method from CAD model is designed, where different assembly relation descriptions in CAD model are summarized into two fundamental relations that are colinear and coplanar. The relation similarity, distance similarity and rotation similarity are adopted as the similar part matching criterions between the CAD model and the observation. The knowledge of part in CAD is used to correct that of the corresponding part in observation. The likelihood maximization estimation is used to infer the accurate poses and assembly relations based on the probabilistic assembly graph.,In the experiments, both simulated data and real-world data are applied to evaluate the performance of the PAGC* model. The experimental results show the superiority of PAGC* in accuracy compared with assembly graph (AG) and probabilistic assembly graph without CAD model (PAG).,The paper provides a new approach to get the accurate pose of each part in demonstration stage of the robot PbD system. By integrating information from visual observation with prior knowledge from CAD model, PAGC* ensures the success in execution stage of the PbD system.

中文翻译：

---

视觉观察与CAD模型相结合的装配任务演示空间信息推理方法

在小零件装配任务的演示（PbD）机器人编程中，在演示阶段通过基于视觉的技术估计的零件姿态的准确性远远不足以确保成功执行。本文旨在开发一种推理方法，通过将视觉观察与计算机辅助设计 (CAD) 模型相结合，提高零件之间的姿态和装配关系的准确性。在本文中，作者提出了一种称为概率装配图的空间信息推理方法。使用可选的 CAD 模型（简称 PAGC*）来完成此任务。然后设计了一种从CAD模型中提取装配关系的方法，将CAD模型中不同的装配关系描述归纳为共线和共面两种基本关系。关系相似度，CAD模型与观测的相似部位匹配准则采用距离相似度和旋转相似度。CAD中零件的知识用于修正观察中相应零件的知识。似然最大化估计用于基于概率装配图推断准确的姿势和装配关系。在实验中，模拟数据和真实世界数据都被用于评估 PAGC* 模型的性能。实验结果表明，PAGC*与装配图（AG）和无CAD模型的概率装配图（PAG）相比，在精度上具有优越性。本文提供了一种在机器人演示阶段获取各部件准确位姿的新方法。 PbD 系统。