단위 행동 학습을 위한 자동 분할 및 확률적 행동유발성 학습

Abstract

In robotics, it is a great challenge for a robot to achieve novel tasks that have not been learnt. The process of learning a human language provides the inspiration that a robot can resolve such a challenge. In the process of learning a language, humans usually recognize where one word ends and another begins (i.e., word boundaries) from the stream of sounds, after which they learn the meanings of the words. Humans then create a variety of novel sentences by combining these words according to their meanings. Just as humans create novel sentences using the learned words, an intelligent robot learns motion primitives embedded in a task by recognizing the boundaries of the motion primitives from motion trajectories demonstrated by a human. After this, the meaning of each motion primitive is learned in order to consider the configurations of task-relevant entities at the boundaries. Ultimately, novel tasks are achieved by recombining the motion primitives in accordance with their meanings.

In this context, intelligent robots should possess three key abilities: i) The ability to recognize the boundaries of the motion primitives. To this end, an autonomous segmentation framework first estimates the boundaries of the motion primitives (i.e., segmentation points) based on Gaussian mixture models (GMMs). To obtain accurate motion primitives, moreover, the GMMs are iteratively remodeled based on their temporal and spatial entropies. Here, the motion primitives should be learned in order to guarantee the achievement of goals under a dynamic environment. For this, the motion primitives are formalized as dynamic movement primitives (DMPs) based on the segmentation points. ii) The ability to learn the meanings of the motion primitives. Hence, the relationships between the DMPs and the information of task-relevant entities are learned as probabilistic affordances. To this end, the training data are first clustered based on effect equivalence, after which the probabilistic affordances are modeled as Bayesian networks with respect to the clusters. iii) The ability to generate the sequence of motion primitives by selecting dependable motion primitives under a dynamic environment. To this end, the probabilistic affordances are arranged according to the sequential structures of given tasks. Ultimately, a dependable motion primitive is selected based on the probabilistic affordances and a motivation value propagation algorithm.

To validate these abilities, various experimental results are presented using six tasks: `cooking rice', `cutting a food item', `setting a table', `painting an assembly part', `assembling a part', and `servicing tea'. |로봇공학에서 로봇이 배운 적이 없는 새로운 임무를 성취할 수 있는 것은 하나의 커다란 도전 과제이다. 사람이 언어를 배우는 과정은 로봇이 그러한 도전 과제를 해결할 수 있는 영감을 제공한다. 사람은 언어를 배우는 과정에서 대개 연속적인 사운드로부터 단어의 시작과 끝을 인식함으로써 단어를 구분한 후 그 단어들의 의미를 학습한다. 그 다음 사람은 그 단어의 의미에 따라 학습된 단어를 재조합 함으로써 수많은 새로운 문장들을 생성한다. 사람이 학습된 단어를 사용하여 새로운 문장을 만들어내는 것과 같이 지능 로봇은 사람의 시연을 통해 획득한 행동 경로로부터 단위 행동의 경계를 찾아냄으로써 임무에 내포되어 있는 단위 행동들을 구분한다. 그런 후에 임무와 관련된 엔티티들의 정보를 이용하여 단위 행동들의 의미를 학습한다. 궁극적으로 새로운 임무들은 단위 행동들의 의미에 따라 그것들을 재조합 함으로써 성취된다.

이런 관점에서 로봇은 세 가지 중요한 능력들을 갖추어야 한다: i) 단위 행동의 경계를 인식하는 능력. 이를 위해 자동 분할 프레임워크가 가우시안 혼합 모델을 기반으로 단위 행동의 경계 (즉, 분할점)를 추정하기 위해 제안된다. 더욱이 정확한 단위 행동들을 획득하기 위해 가우시안 혼합 모델들은 시간과 공간 엔트로피를 기반으로 반복적으로 리모델링 된다. 여기서 단위 행동들은 동적 환경에서 목표 달성을 보장하기 위해 학습되어야 한다. 이를 위해 단위 행동들은 분할점들을 기반으로 동적 단위 행동 (dynamic movement primitives)들로 표현된다. ii) 단위 행동의 의미를 학습하는 능력. 이런 이유에서 동적 단위 행동들과 임무와 관련된 엔티티들 사이의 관계가 확률적 행동유발성으로 학습된다. 이를 위해 학습 데이터들이 우선 결과 동질성 (effect equivalence)을 기반으로 클러스터링 된다. 그런 후에 확률적 행동유발성들은 각 클러스터의 데이터를 이용하여 베이지안 네트워크로 모델링 된다. iii) 동적 환경에서 신뢰할 수 있는 단위 행동을 선택함으로써 단위 행동들의 순서를 생성하는 능력. 이를 위해 확률적 행동유발성들은 임무의 구조적인 성질에 따라 배치된다. 궁극적으로 신뢰할 수 있는 단위 행동은 확률적 행동유발성과 동기 값 전파 알고리즘 (motivation value propagation algorithm)을 기반으로 선택된다.

이러한 지능 로봇이 갖추어야 할 능력들을 검증하기 위해, `밥 짓기’, `음식물 자르기’, `테이블 세팅하기’, `조립 부품에 칠하기’, `조립품 조립하기’, `차 서비스하기’의 여섯 가지 로봇 임무들에 대한 실험 결과를 기술한다.; In robotics, it is a great challenge for a robot to achieve novel tasks that have not been learnt. The process of learning a human language provides the inspiration that a robot can resolve such a challenge. In the process of learning a language, humans usually recognize where one word ends and another begins (i.e., word boundaries) from the stream of sounds, after which they learn the meanings of the words. Humans then create a variety of novel sentences by combining these words according to their meanings. Just as humans create novel sentences using the learned words, an intelligent robot learns motion primitives embedded in a task by recognizing the boundaries of the motion primitives from motion trajectories demonstrated by a human. After this, the meaning of each motion primitive is learned in order to consider the configurations of task-relevant entities at the boundaries. Ultimately, novel tasks are achieved by recombining the motion primitives in accordance with their meanings.

In this context, intelligent robots should possess three key abilities: i) The ability to recognize the boundaries of the motion primitives. To this end, an autonomous segmentation framework first estimates the boundaries of the motion primitives (i.e., segmentation points) based on Gaussian mixture models (GMMs). To obtain accurate motion primitives, moreover, the GMMs are iteratively remodeled based on their temporal and spatial entropies. Here, the motion primitives should be learned in order to guarantee the achievement of goals under a dynamic environment. For this, the motion primitives are formalized as dynamic movement primitives (DMPs) based on the segmentation points. ii) The ability to learn the meanings of the motion primitives. Hence, the relationships between the DMPs and the information of task-relevant entities are learned as probabilistic affordances. To this end, the training data are first clustered based on effect equivalence, after which the probabilistic affordances are modeled as Bayesian networks with respect to the clusters. iii) The ability to generate the sequence of motion primitives by selecting dependable motion primitives under a dynamic environment. To this end, the probabilistic affordances are arranged according to the sequential structures of given tasks. Ultimately, a dependable motion primitive is selected based on the probabilistic affordances and a motivation value propagation algorithm.

To validate these abilities, various experimental results are presented using six tasks: `cooking rice', `cutting a food item', `setting a table', `painting an assembly part', `assembling a part', and `servicing tea'.