Walking Gait of Four-Legged Robot Obtained Throug Q Learning

Walking Gait of Four-Legged Robot Obtained Throug Q Learning

článek ve sborníku ve WoS nebo Scopus

cs

The possible method of walking policy obtaining of four-legged robot through Q-learning is discussed in the contribution. Q-learning is implemented using architecture represented by nondeterministic state machine that defines both possible discrete states and admissible transitions between them. Discrete state is designed as indicators vector of goals achievement by single simultaneously activated instances of two basic controllers. Only simultaneous activations that guarantee static stability of robot are admissible even in the case when single activations cold not achieve its goals. The controllers attempt to achieve its goals using on-line minimization process. Q-learning sequentially improves an estimation of future benefit from usage of admissible simultaneous activations in single discrete states. Walking policy is generated through activations with the highest estimation of future benefit.

The possible method of walking policy obtaining of four-legged robot through Q-learning is discussed in the contribution. Q-learning is implemented using architecture represented by nondeterministic state machine that defines both possible discrete states and admissible transitions between them. Discrete state is designed as indicators vector of goals achievement by single simultaneously activated instances of two basic controllers. Only simultaneous activations that guarantee static stability of robot are admissible even in the case when single activations cold not achieve its goals. The controllers attempt to achieve its goals using on-line minimization process. Q-learning sequentially improves an estimation of future benefit from usage of admissible simultaneous activations in single discrete states. Walking policy is generated through activations with the highest estimation of future benefit.

The possible method of walking policy obtaining of four-legged robot through Q-learning is discussed in the contribution. Q-learning is implemented using architecture represented by nondeterministic state machine that defines both possible discrete states and admissible transitions between them. Discrete state is designed as indicators vector of goals achievement by single simultaneously activated instances of two basic controllers. Only simultaneous activations that guarantee static stability of robot are admissible even in the case when single activations cold not achieve its goals. The controllers attempt to achieve its goals using on-line minimization process. Q-learning sequentially improves an estimation of future benefit from usage of admissible simultaneous activations in single discrete states. Walking policy is generated through activations with the highest estimation of future benefit.

mechatronics, walking robots, machine learning, Q-learning

2003

12.05.2003

Institute of Theoretical and Applied Mechanics, Academy of Sciences of the Czech Republic, Prague

Praha

80-86246-18-3

Engeneering Mechanics 2003

1

2