Method for learning task skill and robot using thereof
View Patent ↗Provided are a method for learning task skill and a robot using the same. The modeling method for learning a task skill includes: receiving training data for a task to be performed by a learning engine; dividing, by the learning engine, the received training data into segments by using a geometric property of a predetermined probabilistic model; and learning, by the learning engine, a basis skill for the divided segments by modeling each divided segment.
1. A modeling method for learning a task skill, which is performed by a learning engine having at least one processor, the modeling method comprising:
receiving training data for a task to be performed by the learning engine;
dividing, by the learning engine, the received training data into segments by using a geometric property of a predetermined probabilistic model; and
learning, by the learning engine, a basis skill for the divided segments by modeling each divided segment;
wherein said dividing of the received training data into segments includes:
determining a segmentation point by analyzing geometric properties of a Gaussian mixture model (GMM) and a Gaussian mixture regression (GMR) for the received training data; and
dividing the training data into segments based on the determined segmentation point.
2. The modeling method for learning a task skill according to claim 1 , wherein said dividing of the received training data into segments calculates variance information of the received training data and determines a segmentation point, which is a criterion of segmentation, based on the calculated variance information.
3. The modeling method for learning a task skill according to claim 1 , wherein said determining of a segmentation point includes:
converting the segmentation point into a dimension, which ensures easy extraction, by reducing the number of dimensions of the training data by means of a principal component analysis (PCA);
generating a Gaussian mixture model by using the converted training data and extracting a window, which is a region where mixture components of the generated Gaussian mixture model overlap each other in time, as a candidate region of a segmentation point; and
selecting a point, where the covariance calculated by the Gaussian mixture regression has a minimal value in the extracted candidate region, as the segmentation point.
4. The modeling method for learning a task skill according to claim 1 , wherein said learning of a basis skill is performed by generating a hidden Markov model (HMM) by using the divided segment.
5. The modeling method for learning a task skill according to claim 1 , further comprising:
receiving new training data for a task to be performed by the learning engine;
generating, by the learning engine, a classifier having a threshold model for classifying the basis skill and the new training data into segments; and
modeling, by the learning engine, a task skill of multi paths expanded from the basis skill by combining the learned basis skills with the new training data classified into segments by using the generated classifier.
6. The modeling method for learning a task skill according to claim 5 , wherein said modeling of a task skill of multi paths generates a model of a single path or multi paths with a divergence by combining the learned basis skills with the new training data classified into segments based on a Kullback-Leibler (KL) divergence.
7. The modeling method for learning a task skill according to claim 5 , wherein said modeling of a task skill of multi paths includes:
calculating a Kullback-Leibler divergence between the model for the learned basis skill and the model for the new training data and a threshold value for the Kullback-Leibler divergence;
comparing the Kullback-Leibler divergence with the threshold value; and
as a result of the comparison, in a case where the Kullback-Leibler divergence is smaller than the threshold value, generating a model having a single path by incorporating the model for the new training data into the model for the basis skill, and in a case where the Kullback-Leibler divergence is equal to or greater than the threshold value, generating a model having a divergence by adding the model for the new training data to the model for the basis skill.
8. The modeling method for learning a task skill according to claim 7 , wherein the threshold value is set to be a minimal value among all statuses between the model for the basis skill and the model for the new training data.
9. A modeling method for learning a task skill, which is performed by a learning engine having at least one processor, the modeling method comprising:
receiving training data for a task to be performed by the learning engine;
dividing, by the learning engine, the received training data into segments by using a geometric property of a predetermined probabilistic model; and
learning, by the learning engine, a basis skill for the divided segments by modeling each divided segment;
wherein said learning of a basis skill is performed by generating a hidden Markov model (HMM) by using the divided segment.
10. A modeling method for learning a task skill, which is performed by a learning engine having at least one processor, the modeling method comprising:
receiving training data for a task to be performed by the learning engine;
dividing, by the learning engine, the received training data into segments by using a geometric property of a predetermined probabilistic model;
learning, by the learning engine, a basis skill for the divided segments by modeling each divided segment;
receiving new training data for a task to be performed by the learning engine;
generating, by the learning engine, a classifier having a threshold model for classifying the basis skill and the new training data into segments; and
modeling, by the learning engine, a task skill of multi paths expanded from the basis skill by combining the learned basis skills with the new training data classified into segments by using the generated classifier.
11. The modeling method for learning a task skill according to claim 10 , wherein said modeling of a task skill of multi paths generates a model of a single path or multi paths with a divergence by combining the learned basis skills with the new training data classified into segments based on a Kullback-Leibler (KL) divergence.
12. The modeling method for learning a task skill according to claim 10 , wherein said modeling of a task skill of multi paths includes:
calculating a Kullback-Leibler divergence between the model for the learned basis skill and the model for the new training data and a threshold value for the Kullback-Leibler divergence;
comparing the Kullback-Leibler divergence with the threshold value; and
as a result of the comparison, in a case where the Kullback-Leibler divergence is smaller than the threshold value, generating a model having a single path by incorporating the model for the new training data into the model for the basis skill, and in a case where the Kullback-Leibler divergence is equal to or greater than the threshold value, generating a model having a divergence by adding the model for the new training data to the model for the basis skill.
13. The modeling method for learning a task skill according to claim 12 , wherein the threshold value is set to be a minimal value among all statuses between the model for the basis skill and the model for the new training data.