Towards algorithmic consideration in robot design: A preliminary study in legged robot

This paper is our first attempt to answer the question: how many legs should a legged robot have, so that the computational cost of full-body motion planning for such a robot can be minimized? To this end, we propose a multi-arm bandit approach to alternate the generation of multiple Rapidly-exploring Random Trees (RRT), where each RRT corresponds to a legged robot with a specific number of legs. The number of legs that correspond to the RRT that finds a valid path the earliest become the preferred number of legs. We also propose a new sampling strategy to speed up the individual RRT-based planning for legged robot. We test the proposed method to find a suitable number of legs (between 4 to 8) for legged robots operating in environments with varying αβexpansiveness, through use of a 'narrow passage' test