Y. Zennir, P. Couturier, and M. Bétemps (France)
Reinforcement learning, Distribution systems, Q-learning, Walking robot
Hexapod robot offers interesting possibilities for studying walking robots. We present the principles of the reinforcement learning used for the training of the walk of a hexapod. The originality of the approach lies in the fact that the training is distributed, each leg has to achieve its own goal. A gait appears as an emerging behaviour and results from the 'self-coordination' of the movements of the legs. This paper describes some simulations that have been carried out in order to study the influence of several training parameters and the comparison of the distributed and centralized approach. We focus on the influence of the choice of the input signals on the training process, and discuss some fault tolerance aspects, thus opening prospects for future works.
Important Links:
Go Back
