Jyotika Bahuguna( Robotics Research Lab, IIIT-Hyderabad)
MS Thesis and Research
My thesis work involved solving the problem of Active Localization for multiple robots by co-operation. For all mobile robotic applications, robot needs to find out where it is in the map. This is called localization. When the robot actively tries to move to places in the map such that localization is faster, the process is called Active Localization. In maps with identical features a robot can have multiple hypotheses about its position. To reduce these number of hypotheses (to one, hence localize), the robot can move to an unique feature in the map. In a multi-robotic scenario, the presence of other robots can help the localization further by the means of robot-robot detection.
This problem requires storing and making sense out of temporal data in the relative sense(i.e, if the robot is in a certain part of the map, what is the probability of it finding an another robot ?). A learning approach to the above problem involves an MDP (Markovian Decision Process). The difference in entropy (about robot's position) is used as the rewards for a direction in which the robot moves. The method ultimately defines a goodness value for each state such that robots move to states/places where the probability of detecting other robots is higher or/and there are unique features of the map (hence localize faster). The role of learning was to make the algorithm more intuitive in localization context.
The method performed well and it was observed that the robots first moved towards areas in the map where they remembered detecting other robots and when that didnt localize them, they moved towards the areas in the map with unique features (Intuitive because, robot-robot detection usually leads to instant localization).The system is hence self-sufficient in the sense that it can localize as efficiently in the presence of other robots as well as in their absence. The direct advantage of this behavior is that the system need not be trained everytime a robot is added to the group. An another trivial observation was that, negative entropy as a feedback in markov models lead to more stabler systems by avoiding them from getting stuck in a loop of never progressing state transitions.