Virtual Model Generation for Robust Robot Control Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training control modules for operating machines, such as industrial robots, using simulation face inefficiencies due to the generation of impossible virtual models, leading to reduced learning efficiency and inability to apply learned control modules to real-world variations in components and environments.
Innovation Solution
An information processing device that stochastically generates virtual models based on input parameters and probability distributions for characteristics of constituent components, allowing for successful operation determination and control module generation robust against real-world variations, thereby reducing learning time and improving adaptability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If learning is performed by randomly generating virtual models, then the control module can be obtained through simulation, but actually impossible virtual models are also learned which deteriorates learning efficiency
Solution Approach 1:
The patent changes the parameter generation method from completely random to probability distribution-based. By defining specific probability distributions for component characteristics (e.g., robot arm length, sensor position) that reflect real-world constraints, the system generates virtual models with realistic variations while excluding impossible configurations, thus improving both learning efficiency and control module quality.
Solution Approach 2:
The patent performs preliminary action by pre-defining probability distributions that encode real-world physical constraints before the learning process begins. This preliminary setup ensures that all generated virtual models are physically plausible, preventing the learning process from wasting time on impossible configurations while still providing sufficient variation for robust control module development.
2Productivity
If only actually possible virtual models are learned, then learning efficiency is improved, but the obtained control module cannot be applied to unlearned and unknown situations
Solution Approach 1:
The patent uses probability distributions to generate virtual models with realistic variations in component characteristics. This approach allows the system to explore the full range of physically possible configurations (improving adaptability) while maintaining learning efficiency by excluding impossible models. The controlled randomness enables the control module to handle unlearned situations within real-world constraints.
3Reliability
If virtual models with variation in component characteristics are used, then the control module becomes robust against real-world variation, but the complexity of model generation increases
Solution Approach 1:
The patent manages complexity by using probability distributions to systematically vary component characteristics. This mathematical approach provides a structured method for introducing realistic variations (improving robustness) while maintaining manageable complexity through well-defined statistical parameters rather than ad-hoc model generation.
Solution Approach 2:
The probability distribution framework serves multiple functions: it generates realistic variations, ensures physical plausibility, controls the range of variations, and simplifies the generation process. This universal approach handles both the robustness requirement and the complexity management in a unified manner.
Data Source
AI summary
The image processing device is provided with: a first input unit which, with respect to one or more virtual models including a virtual model of an operation machine, receives an input of a first parameter for identifying a type; a second input unit which receives an input of a second parameter relating to a stochastic distribution having, as a random variable, a characteristic of an element constituting the one or more virtual models; a virtual model generation unit which, using the first parameter and the second parameter, generates the one or more virtual model stochastically; a determination unit which determines the correctness of an operation of the virtual model of the operation machine when operated in a virtual space including the one or more stochastically generated virtual models; and a learning unit which learns a control module for the operation machine for achieving a predetermined operation.