Machine Learning Hand Selection for Robot Grasping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for selecting a hand to transfer workpieces by robots are limited, as they require pre-taught form data for industrial products with fixed shapes, making it difficult to handle varying shapes and sizes of non-industrial items like vegetables or fruits, and managing multiple types of industrial products.
Innovation Solution
A machine learning device that uses image data and identification or physical information of successful hand interactions to construct a learning model, enabling the selection of appropriate hands for transferring workpieces without prior teaching, by employing state observation, label acquisition, and supervised learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If form data of workpieces is taught in advance in association with hands, then hand selection is automated for fixed-shape industrial products, but the system cannot handle workpieces with varying shapes and sizes such as vegetables and fruits
Solution Approach 1:
The patent replaces the mechanical system of pre-taught form data associations with a machine learning-based image recognition system. The learning model processes images of workpieces and automatically determines appropriate hands without requiring manual form data entry or association, thereby handling varying shapes and sizes effectively.
Solution Approach 2:
The system enables self-service by allowing the machine learning model to automatically learn and determine hand selections based on image data. The system serves itself by continuously improving its hand selection capability through learning from training data, eliminating the need for manual programming of form data associations.
2Adaptability or versatility
If form data is prepared for all types of industrial products, then comprehensive hand selection is possible, but it becomes difficult to manage form data for a large number of product types
Solution Approach 1:
The patent substitutes the manual mechanical process of preparing and managing form data with an automated machine learning system. The learning model automatically processes images and determines hand selections, eliminating the time-consuming manual form data preparation while maintaining comprehensive coverage across multiple product types.
Solution Approach 2:
The system uses image copies of workpieces as training data for the machine learning model. Instead of manually creating form data for each product type, the system learns from visual copies (images) of various workpieces, automatically generalizing to handle diverse product types without explicit form data preparation.
3Reliability
If pre-taught combinations of workpieces and hands are used, then hand selection is straightforward for known products, but the system fails when encountering unknown or untaught workpiece types
Solution Approach 1:
The patent introduces dynamics by using a machine learning model that can adapt and learn from new data. Unlike static pre-taught combinations, the learning model continuously improves its ability to select appropriate hands for both known and unknown workpiece types by learning from training data and making predictions based on image features.
Solution Approach 2:
The system changes parameters by using image data features (such as shape, size, and visual characteristics) instead of relying on pre-defined form data categories. The machine learning model learns to map image parameters to appropriate hand selections, enabling reliable handling of both familiar and novel workpiece types through parameter-based generalization.
Data Source
AI summary
A hand for transferring a transfer target can be selected even when the combination of the transfer target and the hand is not taught. A machine learning device includes: state observation means for acquiring at least a portion of image data obtained by imaging a transfer target as input data; label acquisition means for acquiring information related to grasping means attached to a robot for transferring the transfer target as a label; and learning means for performing supervised learning using a set of the input data acquired by the state observation means and the label acquired by the label acquisition means as teacher data to construct a learning model that outputs information related to the grasping means appropriate for the transferring.