Interactive Learning from Demonstration with a Multilevel Mechanism for Collision-Free Navigation in Dynamic Environments

While collision-free navigation could be accomplished using existing rule-based approaches, it would be more attractive to use learning from demonstration (LfD) approaches to ease the burden of tedious rule designing and parameter tuning procedures. To increase the performance interactively, the basic LfD method is further combined with the interactive learning approach in this paper. However, instead of retraining the control policy using all available data, we propose an interactive learning approach with a multilevel mechanism to learn the control policy interactively for accomplishing collision-free navigation in dynamic scenes. Compare with the approach in which all the available data are used for retraining, not only the collision rate decreases, the training time also decreases since the top level policy is not retrained and the training time for the additional level is much less than retraining the whole policy. The results show that collision-free navigation in simulated dynamic environments is learnable from interactive demonstrations.