Risk Manager

The goal of this agent is to keep the risk of losing the target as low as possible. While the Target Tracker's goal is to locate the target by maintaining it in the camera's view field, this agent tries to keep a reasonable amount of known landmarks, as non collinear as possible, in the surroundings of the robot. The rationale is to have as many visible landmarks as possible so that the Map Manager is able to compute the location of the target using the beta-coefficient system when it is not visible nor in the Visual Memory. The fewer surrounding landmarks whose locations are known, the more risky is the current situation and the higher the probability of losing the target and getting lost. Also, the more collinear the landmarks, the higher the error in the location of the target, and thus, the higher the imprecision on its location.

We model the risk as a function that combines: 1) the number of landmarks ahead (elements in set $\mathit{A}$), 2) the number of landmarks around (elements in set $\mathit{B}$), and 3) their ``collinearity quality'' ($q_{A}$ and $q_{B}$). As we have described, these qualities are computed by the Map Manager. A minimum risk of 0 is assessed when there are at least six visible landmarks in the direction of the movement and minimally collinear. Although the locations of only three landmarks are needed in order to use the beta-coefficient system, we want to have additional landmarks around the robot whose locations are known, so that there are more chances to compute the target's location. A maximum risk of 1 is assessed when there are no landmarks ahead nor around:

$$R=1 - \min \left(1, q_{A} {\left(\frac{\vert A \vert}{6}\ri... ...}{\left(\frac{\vert B \vert}{6}\right)}^{\gamma_{B}}\right)$$ (9)

The values $\gamma_{A}$ and $\gamma_{B}$ determine the relative importance of the situation of landmarks (ahead or around).

Given that the robot cannot decrease the collinearity of the visible landmarks, the only way to decrease the risk level is by increasing the number of landmarks ahead and around. Having more landmarks, besides increasing $\vert A \vert$ or $\vert B \vert$, also helps by possibly increasing the qualities $q_A$ and $q_B$.

We encourage having landmarks ahead by bidding

$$bid\left(look\left(random\left(\left[-\frac{\pi}{4},+\frac{\pi}{4}\right]\right)\right)\right) = \gamma_r \cdot R$$ (10)

for the action of looking at a random direction in front of the robot and trying to identify the landmarks in that area, if $\vert A \vert < 6$, and

$$bid\left(look\left(random\left(\left[+\frac{\pi}{4},+\frac{7\pi}{4}\right]\right)\right)\right) = \gamma_r \cdot R^{2}$$ (11)

(which is obviously smaller than $\gamma_r \cdot R$) for the action of looking at a random direction around the robot and trying to identify landmarks, if $\vert B \vert < 6$, where $\gamma_r$ is a parameter to control the maximum value of the bidding function. The bidding functions are shown in Figure 4.6.

Figure 4.6: Risk Manager's look bidding functions (look ahead -solid line- and look behind -dashed line-)

The behavior of this agent also helps the Map Manager build the map when the robot is in an unexplored area. Since it bids for looking for landmarks when there are not many visible, its bids will be high, and thus new landmarks (if there are landmarks, obviously) will be identified and the map will be updated.