Robot Holding Mode Inference Using Hierarchical Image Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training robot hands to hold objects effectively lack accuracy in inferring the optimal holding mode, which affects the robot's ability to grasp and manipulate objects correctly.
Innovation Solution
A trained model generation method that includes a class inference model, a first holding mode inference model, and a second holding mode inference model, which classify and infer the holding category and modes for a robot to grasp objects based on images of the objects, using learning data to improve the accuracy of holding mode determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single holding mode inference model is used, then the device complexity is low, but the measurement precision of holding mode inference is insufficient
Solution Approach 1:
The patent divides the holding mode inference task into multiple specialized models: a class inference model for object classification, a first holding mode inference model for initial holding mode prediction, and a second holding mode inference model for refined holding mode determination. Each model focuses on specific aspects of the inference task, improving overall accuracy while maintaining manageable complexity through functional segmentation.
Solution Approach 2:
The patent introduces a hierarchical dimension to the inference process by processing data through multiple stages with different levels of abstraction. The class inference model operates at the object level, the first holding mode inference model operates at the category level, and the second holding mode inference model operates at the specific action level, adding dimensional depth to the inference architecture.
2Measurement precision
If multiple inference models are used to improve accuracy, then the measurement precision increases, but the loss of time for processing increases
Solution Approach 1:
The class inference model performs preliminary classification of the object before the holding mode inference models process the data. This preliminary action filters and categorizes input data early in the pipeline, reducing the complexity of subsequent processing and enabling faster execution of the detailed holding mode inference while maintaining high accuracy.
3Reliability
If simple learning data is used, then the ease of manufacture of the trained model is high, but the reliability of the inference result is low
Solution Approach 1:
The patent applies different training data requirements and processing methods to different components of the multi-model system. Each inference model is trained with learning data optimized for its specific function, allowing high reliability to be achieved through specialized, targeted training rather than uniformly complex training across all models.
Data Source
AI summary
A trained model includes a class inference model, a first holding mode inference model, and a second holding mode inference model. The class inference model is configured to infer, based on an inference image of a holding target object, a classification result obtained by classifying the holding target object into a predetermined holding category. The first holding mode inference model is configured to infer, based on the classification result and the inference image, a first holding mode for the holding target object. The second holding mode inference model is configured to infer, based on the first holding mode and the inference image, a second holding mode for the holding target object. A trained model generation method includes generating a trained model by performing learning using learning data including a learning image of a learning target object corresponding to a holding target object.