This paper is available on arxiv under CC BY-NC-ND 4.0 DEED license.
Authors:
(1) Akash Chikhalikar, Graduate School of Engineering, Department of Robotics, Tohoku University;
(2) Ankit A. Ravankar, Graduate School of Engineering, Department of Robotics, Tohoku University;
(3) Jose Victorio Salazar Luces, Graduate School of Engineering, Department of Robotics, Tohoku University;
(4) Yasuhisa Hirata, Graduate School of Engineering, Department of Robotics, Tohoku University.
In this section, we elaborate on our multi-target search method. We consider the task of locating two targets, cup and remote. However, our method can be easily extended to any number of objects. We first extract a quantitative relationship between the targets to be found and the landmark locations. Next, we describe our methodology for the heuristic search.
A. Probabilistic Priors
Priors help determine the probability for finding an target near a known associated landmark. The relationship between two associated objects is defined by ontology. In our scenario, we interpret the ontology isNear(TargetA, LandmarkB) as the probability of finding Target ‘A’ and Landmark ‘B’, in the same image frame. The location of Target ‘A’ is unknown while the location of Landmark ‘B’ is known within a certain range. The associated probabilistic priors are calculated from the publicly available data set ‘Common Objects in Context (COCO) [29]. Figure 3 shows the normalized probabilistic relationships of finding Target ‘A’ near Landmark ‘B’.
B. Heuristic search
Heuristics are essential for evaluating the cost of visiting a location to search for objects. They should minimize the energy spent while maximizing the chances of finding the object. Furthermore, they need to consider both targets and incorporate user preferences, especially under adversarial conditions. We consider two heuristics for performing a multi-object search. The first heuristic acts as the baseline and is defined as follows:
This paper proposes a novel heuristic for a multi-target
search that includes user preference:
The hyperparameters α and β are used to incorporate user preferences. If the user prioritizes target 1, then α > β, and vice versa. The distance to visit the location l is obtained from the A* global planner. A greedy search is initiated according to this heuristic. When one of the objects is found, the heuristic drops down to its single-object equivalent. The search ends when both targets are found. The following pseudocode is an overview of our object-search and navigation strategy.
The user-defined priority also changes the trajectory taken by the robot while searching for targets. In Fig. 4, the robot follows the sequence, Table →Sofa→TV when searching for cup is the priority. If the user prioritizes remote, then the sequence of locations visited is Sofa →TV→Table. For an equal priority search, the robot follows Sofa→Table→TV sequence.