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.
Table of Links
- Abstract & Introduction
- Related Work
- Semantic Mapping Framework
- Multi-Target Search
- Experimental Studies
- Conclusions and Discussions, Acknowledgment and References
II. RELATED WORK
Previously, researchers have tried to solve the target search/retrieval problem in three ways. One approach is to formulate target search as a Partially Observable Markov Decision Process (POMDP) and optimize the semantic gain from sensor observations at every time step. The Informative Path Planning (IPP) approach aims to develop an informative costmap corresponding to the metric map to supplement target search. Lastly, the Next Best View (NBV) methods consider target search to be a discrete problem and use different strategies to determine best locations for search. A similar area of research is that of navigating to an object in an unseen environment. Researchers working on this problem make use of photo-realistic simulators which have been introduced in the Habitat challenges [11], [12].
A. POMDPs
POMDPs are a generalization of MDPs by including uncertainties in the observation along with state transition uncertainty. These have been popular among researchers trying to push the state of-the-art in this domain. However, POMDPs suffer from intractability when solving for large domains [13]. To tackle the computational burden, researchers have introduced hierarchies in spatial scales or planning [14]–[16].
B. Informative Path Planning (IPP)
In this method, path planning is driven by a joint cost function consisting of information gained and distance traveled along the path. [17] first uses Gaussian Mixture Models and bayesian relationships to prepare a information map. A sampling-based IPP is prepared using this map for object search. IPP can also benefit from Reinforcement Learning, as explored in [18].
C. Next Best View (NBV)
NBV methods, as described earlier, depend on determining best locations for target search. In [19], the authors evaluate all routes to find the target object and then store information about different objects seen during navigation for quicker retrieval in the next task. The authors in [20] determine a set of candidate view points and evaluate the view points based on probabilistic belief around the view point. Our approach can be broadly categorised as a Next Best View method. Our novelty exists in extending previous approaches to include multiple targets as well as introduce capabilities to prioritise targets individually.
