6D Assembly Pose Estimation by Point Cloud Registration for Robot Manipulation
(2312.02593)Abstract
The demands on robotic manipulation skills to perform challenging tasks have drastically increased in recent times. To perform these tasks with dexterity, robots require perception tools to understand the scene and extract useful information that transforms to robot control inputs. To this end, recent research has introduced various object pose estimation and grasp pose detection methods that yield precise results. Assembly pose estimation is a secondary yet highly desirable skill in robotic assembling as it requires more detailed information on object placement as compared to bin picking and pick-and-place tasks. However, it has been often overlooked in research due to the complexity of integration in an agile framework. To address this issue, we propose an assembly pose estimation method with RGB-D input and 3D CAD models of the associated objects. The framework consists of semantic segmentation of the scene and registering point clouds of local surfaces against target point clouds derived from CAD models to estimate 6D poses. We show that our method can deliver sufficient accuracy for assembling object assemblies using evaluation metrics and demonstrations. The source code and dataset for the work can be found at: https://github.com/KulunuOS/6DAPose
Overview
-
Introduction of a novel method for 6D assembly pose estimation using RGB-D data and 3D CAD models.
-
The research addresses the complexity of robotic assembly, aiming for agile systems within robotic frameworks.
-
The method involves semantic segmentation, projection of point clouds, and an iterative process independent of previous steps.
-
Evaluation through synthetic datasets, showing promising results in estimating accurate assembly poses.
-
Future work may explore learning-based modules and complex tasks like insertion or clamping.
Introduction to Assembly Pose Estimation
Robotic manipulation skills are becoming increasingly sophisticated and are vital for tasks such as industrial manufacturing, mining, and space explorations. To approach robotic assembly, a robot must be able to perceive the environment and determine the precise pose (position and orientation) of components for successful assembly. Assembly pose estimation is the process of determining how to place an object in relation to other parts.
Challenges in Robotic Assembly
In traditional robotic assembly research, the focus has primarily been on initial pose estimation of objects. For a robot to perform assembly tasks that include putting things together according to specific constraints, it requires a more detailed insight into the relative positioning of parts – known as the assembly pose. Research in this area faces issues due to its complexity and the need for such systems to be agile, reacting quickly to changes and integrating seamlessly into robotic frameworks.
The Proposed Solution
A novel method is presented in this research to estimate the 6D assembly pose, utilizing both RGB-D data (color and depth information) and 3D CAD models of objects. This involves:
- The adaptation of established object pose estimation methods for assembly pose estimation.
- The creation of source point clouds using CAD models which are critical for accurate registration.
- An iterative process for estimating assembly poses in multi-object assemblies.
- Evaluation of point cloud registration effectiveness in pose estimation.
This method starts with a semantic segmentation module to identify objects within a scene, followed by projecting point clouds from both the scene and the CAD models. Assembly poses are determined sequentially for each step, assessing them independently instead of relying on the success of previous steps. The method can be integrated with existing pose estimation and grasp detection methods without the need for additional model training.
Evaluation and Results
The research outlines the generation of synthetic datasets specifically for the evaluation of assembly pose estimation, due to the absence of standard datasets for this problem. Two simulated gear assembly datasets were created and evaluated using metrics that consider both the registration accuracy of point clouds and the estimated assembly pose itself. The method was also validated with an industrial Diesel engine assembly use case.
The results demonstrate that accurate 6D assembly poses can be estimated for object assemblies, contributing significantly to the advancements in robotic manipulation tasks. However, the accuracy may be affected by factors such as occlusions or the complexity of the assembly involving multiple objects. Future work could include learning-based pose estimation modules and addressing more complex tasks that involve additional constraints, like insertion or clamping tasks.
In conclusion, this paper provides a step forward in enabling robots to execute precise and intelligent assembly tasks by accurately estimating assembly poses through point cloud registration. This approach has the potential to enhance the efficiency and reliability of robots in various high-skill application areas.
Create an account to read this summary for free: