Robotic Grasp Proposal Generation Without 6D Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object grasping systems struggle with grasping objects that are overlapping, in random poses, deformable, or with varying appearances, as they rely on biased object pose detectors and traditional keypoints, leading to accumulated errors and inefficiencies.
Innovation Solution
The method determines invariant features of objects, such as keypoints and masks, to generate candidate grasp proposals without relying on 6D object pose estimation, using a robotic system with an end effector and computing system to execute grasp proposals effectively across various object configurations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional object pose detectors and traditional keypoints are used for grasping, then the system can operate with simple algorithms, but it fails to accurately grasp overlapping, deformable, or appearance-varying objects due to accumulated errors
Solution Approach 1:
The patent segments the grasping task into distinct components: keypoint detection, mask generation, and grasp proposal generation. By dividing the complex pose estimation problem into these separate stages, the system achieves higher reliability in grasping accuracy without requiring a single complex algorithm to handle all aspects simultaneously.
Solution Approach 2:
The patent performs preliminary actions by first detecting keypoints and generating masks before creating grasp proposals. This preliminary detection and segmentation of the object provides a more reliable foundation for subsequent grasping operations, reducing accumulated errors that would occur if all operations were performed in a single step.
2Loss of information
If 6D object pose estimation is performed to determine grasp locations, then comprehensive object information is obtained, but errors accumulate through multiple intermediate steps reducing grasping success
Solution Approach 1:
The patent extracts only the necessary information (keypoints and masks) required for grasping rather than performing complete 6D pose estimation. By taking out only the essential geometric features needed for grasp proposal generation, the system avoids the error accumulation inherent in multi-step pose estimation while still obtaining sufficient object information for accurate grasping.
Solution Approach 2:
Instead of following the conventional approach of pose estimation first and then grasp proposal, the patent inverts the workflow by directly generating grasp proposals from detected keypoints and masks. This inversion eliminates the intermediate pose estimation step that causes error accumulation, while still achieving comprehensive object information utilization for grasping.
3Measurement precision
If the system processes each object individually through pose detection, then detailed object information is obtained, but the process becomes inefficient for scenes with multiple overlapping objects
Solution Approach 1:
The patent segments the scene processing by detecting keypoints and generating masks for each object individually, then efficiently generating grasp proposals from these segmented representations. This segmentation approach maintains measurement precision for each object while improving overall productivity by avoiding repeated full pose estimation for overlapping objects.
Solution Approach 2:
The patent uses detected keypoints and masks as simplified copies or representations of the full object geometry. These copied representations contain sufficient information for grasp proposal generation, allowing the system to process multiple overlapping objects efficiently without performing complete pose detection on each object, thus maintaining precision while improving productivity.
Data Source
AI summary
A method for object grasping, including: determining features of a scene; determining candidate grasp locations; determining a set of candidate grasp proposals for the candidate grasp locations; optionally modifying a candidate grasp proposal of the set; determining grasp scores associated with the candidate grasp proposals; selecting a set of final grasp proposals based on the grasp scores; and executing a grasp proposal from the set of final grasp proposals.