Layered Cost map Navigation. The main difference with the previous case is the use of continuous state features. From this learning, we build a cost map from which navigation is made possible. − Since the states taken into account correspond to the polar human representation, we set n number of random points in the environment within a range for each axis of rd = [0, 14] and rθ = [ π, π), where r represents range. This draw can be seen as the points in Figure 5(b) and they represent the mean in the 2D gaussian used for the RBF. As for the value of the standard deviation, all RBF bins have the same value which is a quarter of the range for each axis. Thus, the vector state representation is Φ(s) = [φ1(scoord), φ2(scoord), . . . , φn(scoord)], where φi(scoord) is the ith RBF and scoord is the cartesian center of the state s. Then we set Φ(s, a) = Φ(s) given than it is intended to use this information in a cost map, which is only represented by the states and not the actions, differently from Naive Global Planner. For the Layered Cost map methodology, after the learning process the weight vector w is set. One important point is that Φ(s, a) = Φ(s) and s is represented by spatial features. Thus, a cost map can be generated in the environment. Figure 5(b) shows a cost map like result of the demonstrations given in Figure 3, this is feasible due to the representation of features as continuous functions. Even when we have discrete states, the values of the coordinate system is in for distance and angle. The result in the simulated scenario is depicted in Figure 5(a).
