Robot Grasp Pose Prediction From 3D Point Clouds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for determining the grasping posture of a robot gripper are inefficient, particularly in high-dimensional scenarios, and require costly CAD models, while they often only handle objects grasped from above and lack flexibility in grasping postures.
Innovation Solution
A learning device that uses a neural network model to detect object locations and postures through image data from an RGB-D camera, generating supervised data to learn a grasping model capable of predicting optimal gripper positions and orientations for various objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional methods (neural network or deep learning) are used to estimate gripper location and orientation, then the method is widely applicable, but it is difficult to find high-dimensional grasping posture and predict information that is difficult to annotate in images
Solution Approach 1:
The patent transitions from 2D image annotation to 3D point cloud data representation. By using three-dimensional point cloud information instead of two-dimensional image data, the system can accurately predict high-dimensional grasping postures including position, orientation, and gripper configuration, overcoming the limitation of conventional 2D-based methods that struggle with high-dimensional posture prediction.
2Measurement precision
If CAD model recognition method is used to recognize three-dimensional objects, then object recognition is achieved, but it requires high economical and temporal costs because a CAD model is required and the grasping posture has to be determined after the object is recognized
Solution Approach 1:
The patent uses point cloud data as a direct digital copy of the object's three-dimensional structure, eliminating the need for separate CAD model creation and matching processes. The point cloud data obtained from depth sensors directly represents the object geometry, allowing immediate grasping posture determination without the time-consuming CAD model recognition pipeline.
Solution Approach 2:
The patent replaces the mechanical/CAD-based object recognition system with a direct point cloud processing approach using machine learning. Instead of requiring CAD models and geometric matching algorithms, the system uses neural networks to directly predict grasping postures from point cloud data, significantly reducing processing time and computational complexity.
3Ease of operation
If conventional methods are used for grasping, then objects can be grasped from right above, but no examples of learning grasping methods other than that can be found
Solution Approach 1:
The patent enables dynamic and versatile grasping by predicting full 6-degree-of-freedom grasping postures (position and orientation) rather than fixed top-down grasps. The system can adapt to different object shapes, sizes, and orientations, allowing grasping from various angles and positions, making the robot arm more flexible and adaptable to diverse grasping scenarios.
Data Source
AI summary
An estimation device includes a memory and at least one processor. The at least one processor is configured to acquire information regarding a target object. The at least one processor is configured to estimate information regarding a location and a posture of a gripper relating to where the gripper is able to grasp the target object. The estimation is based on an output of a neural model having as an input the information regarding the target object. The estimated information regarding the posture includes information capable of expressing a rotation angle around a plurality of axes.