Applications of Experience Graphs in Humanoid Motion Planning

The capability of using cached plans, whether in the form of expert provided demonstrations or previously generated paths from prior planner runs, is a useful addition to any search based planner. Its benefits range from faster planning times in large state spaces to generation of more predictable replans in case of partially known environments which robots discover as they plan. The experience graph (EGraph) algorithm is a planning algorithm which allows a planner to reuse cached plans/demonstrations by providing a heuristic which biases the search towards states which are present in prior demonstrations. This technical report addresses the particular application of the experience graph algorithm in search based planning for a 35-DOF humanoid. In this work, we outline the details of deploying E-Graphs to this particular problem and show that using E-Graphs significantly outperforms conventional methods like weighted A*.