Adaptive point-based value iteration for continuous states POMDP in goal-directed imitation learning

In motion planning and robot navigation, continuous domain would be the natural way of representation of state space. However, discretization is needed in order to deal with continuous state space. Results precision depends on the discretization, which leads to a problem of “curse of dimensionality”. We present a new approximation approach of goal-directed imitation learning algorithm using the point-based value iteration algorithm that deals with continuous domain in motion planning. We demonstrate our algorithm in the V-REP robot simulator, to validate the experimental result.