Robot Contact Determination Model for Low-Data Teaching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for teaching robot devices to perform tasks with varying manipulator mechanisms and workpieces are costly and inefficient, as they require extensive manual programming and data storage for expressing object interactions in high-dimensional coordinate spaces, leading to increased data volume and storage challenges, especially for systems with limited capacity.
Innovation Solution
A model generation apparatus and control apparatus that utilize machine learning to generate and apply a determination model to predict object contact based on positional relationships, reducing data volume and enabling efficient operation of robot devices in versatile situations without direct association with specific tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If information indicating the boundary of contact between two objects is expressed using coordinate space with high resolution, then the precision of contact detection is improved, but the data volume increases dramatically
Solution Approach 1:
The patent replaces the traditional geometric/mathematical method of expressing contact boundaries in coordinate space with a machine learning-based determination model. Instead of using high-resolution coordinate data to represent contact boundaries, the system uses a trained neural network model that processes object feature information to determine contact relationships, thereby avoiding the exponential data volume increase associated with high-dimensional coordinate representations.
Solution Approach 2:
The patent changes the parameters used for contact detection from high-resolution coordinate space representations to feature-based representations processed by a determination model. By transforming the input parameters from detailed spatial coordinates to extracted object features, the system maintains contact detection precision while significantly reducing the data volume required to represent contact boundaries.
2Manufacturing precision
If manual teaching methods are used to program robot devices for varying components and workpieces, then the robot can perform specific tasks with high precision, but the time and cost for teaching increase significantly
Solution Approach 1:
The patent enables the robot device to perform tasks autonomously by equipping it with a determination model that can independently detect contact relationships between objects. Instead of requiring manual programming for each task scenario, the robot uses the machine learning model to automatically determine contact boundaries and adjust its operations, thereby eliminating the time-consuming manual teaching process while maintaining task execution precision.
Solution Approach 2:
The patent performs preliminary training of the determination model using contact boundary information obtained through manual teaching or other methods. Once the model is trained, it can be reused for multiple tasks and scenarios, eliminating the need for repeated manual teaching. The preliminary action of model training stores the knowledge needed for contact detection, allowing the robot to perform subsequent tasks autonomously without additional teaching time.
Data Source
AI summary
A model generation apparatus according to one or more embodiments may include: a data obtainer configured to obtain a plurality of learning datasets each including a combination of training data and true data, the training data indicating a positional relationship between two objects, the true data indicating whether the two objects come in contact with each other in the positional relationship; and a machine learning unit configured to train, through machine learning, a determination model using the obtained plurality of learning datasets to cause the determination model to output, in response to an input of training data included in each of the plurality of learning datasets, an output value fitting true data included in a corresponding learning dataset of the plurality of learning datasets.