A series of kinodynamic sampling-based planners
have appeared over the last decade to deal with high dimen-
sional problems for robots with realistic motion constraints. Yet,
offline sampling-based planners only work in static and known
environments, suffer from unbounded memory requirements
and the produced paths tend to contain a lot of unnecessary
maneuvers. This paper describes an online replanning algo-
rithm which is flexible and extensible. Our results show that
using a sampling-based planner in a loop, we can guide the
robot to its goal using a low dimensional navigation function.
We obtain higher success rates and shorter solution paths in a
series of problems using only bounded memory.