A Multi-Strategy Architecture for On-Line Learning of Robotic Behaviours

A Multi-Strategy Architecture improves the efficiency of on-line learning of robotic behaviours by taking inspiration from approaches humans use for learning complex behaviours. The hybrid approach first learns the qualitative dynamics of a robotic system from which a symbolic planner constructs an approximate solution to a control problem by qualitatively reasoning over the discovered dynamics. The parameters of the approximate solution are refined by numerical optimization, into a policy for a reactive controller. The hybrid approach is demonstrated on a multi-tracked robot intended for urban search and rescue.