Combination of Reinforcement Learning Models Towards considerate motion planning for multiple pedestrians

Abstract

This paper addresses a human in the loop reinforcement learning(HiL-RL) model to realize compassionate behavior where humans and robots co-exist in a simulation environment. A composition method of RL models is proposed to deal with multiple pedestrian and scalability issues. In HiL-RL a single human pedestrian gives a reward to the robot corresponding to the robots behaviour. To deal with multiple pedestrians and improve scalability, trained RL models for each pedestrian are combined instead of executing Hil-RL with multiple pedestrians. Three combination methods, one where the Q values for a given state and action pair are combined and the two more where the velocities for a given state are combined but using slightly different methods are proposed. The robots behaviour is then determined and compared for each of these models in an environment where two and four pedestrians are present. The robot sucessfully chnages its motion in order to the avoid collision with nearby pedstrians.