POMDP Navigation of Service Robots with Human Motion Prediction

To improve the natural pedestrian-avoidance skills of service robots in indoor dynamic environments,a method of combining long-term and short-term prediction of pedestrian's motion is introduced on the basis of modeling humans'motion trajectory patterns.In order to accommodate the uncertainties in the perception-control loop of robots,which are mainly caused by sensor noise and time delay in network and other factors,the relative position relation between human and robot is modeled as partially observable Markov state.Partially observable Markov decision process(POMDP) is utilized for probabilistic decision-making under multi-source uncertainties,and the behavior modules of the global path planner,the motion reactor and the speed controller,are coordinated.Experimental results illustrate the performance of safe navigation that can avoid conflicts in advance,as well as the improved robot navigation efficiency by avoiding repeated zigzaging and wandering motion.