Burak Kaleci

Market-based task allocation by using assignment problem

In this study, a market-based task allocation method is proposed. In the trading process, the energy model of a robot platform is used to calculate the price and cost for a task. In order to determine the winner robot(s), two auction clearing algorithms, which are named as Iterative and Highest-Bid Task for Robots (HBTR), are proposed. Additionally, assignment problem is used to determine instantaneous optimal task-robot matching. The Hungarian algorithm is implemented to solve optimal assignment problem. In the implementation of the algorithms, three types of tasks are used: cleaning, carrying and monitoring. The tasks consist of three important features: delicacy, priority, and the task completion time. These tasks are assigned to the members of a heterogeneous robot team, according to the proposed task allocation method.

A probabilistic approach for semantic classification using laser range data in indoor environments

In this paper, a probabilistic approach is proposed for semantic classification in indoor environments using laser range data. Robot locations in indoor environments are categorized into three broad classes as room, corridor, and door. K-means and Learning Vector Quantization (LVQ) methods are used to classify robot positions. Circular shifting is applied to render laser range data independent of robot pose. K-means or LVQ algorithms are used to determine data clusters and their centers. In K-means method, the cluster centers are modelled with the proposed probabilistic approach to consider the semantic class of robot location. On the other hand, LVQ method inherently provides semantic classes of the cluster centers. In order to improve the rate of classification success, Markov model is integrated into the proposed approach. Experiments are conducted to demonstrate the effectiveness of the proposed approach. The results indicate that K-means method successfully classifies rooms and corridors, but door classification success rate is not satisfactory. LVQ method improves door classification rate without decreasing the classification rate of corridor and room. Lastly, effectiveness of the Markov model is discussed.

Rule-Based Door Detection Using Laser Range Data in Indoor Environments

In indoor environments, doors generally separate different parts of buildings such as rooms and corridors. The information whether a robot is at a door location is valuable for navigation, localization, mapping, and exploration tasks. Existing algorithms focus on vision-based methods which are supported by laser range data to detect closed doors along a corridor. In this paper, a rule-based door detection method that uses only laser range data is presented. The proposed method is based on three assumptions: Firstly, the doors must be open. Also, the robot should be near and/or between the door frames for proper detection of the door. Lastly, doors are assumed to be along the axes of the reference coordinate frame. Under these assumptions, there is always a bottleneck at a certain location in the shape formed by the laser beam readings. Our method exploits this observation to define a set of rules for detecting doors. Simulations are performed using Freiburg 79 and ESOGU Electrical Engineering Laboratory buildings data to measure the performance of the proposed algorithm. The algorithm detected 90% of the doors for both environments with a false positive rate smaller than 6.1% and 0.7% for the Freiburg 79 and ESOGU buildings, respectively.

A new sensor model for particle-filter based localization in the partially unknown environments

Localization is an important ability for a mobile robot. The probabilistic localization method becomes more popular because of the ability of representing the uncertainties of the sensor measurements and inaccuracy environments, robust solutions for a wide perspective of localization problem. The particle filter is one of the Bayesian-based methods. In this study, a new sensor model (R2SM) is proposed and integrated to traditional particle filter to reduce the effects of outliers. The proposed approach is applied in the global and position tracking localization problems for a mobile robot in static and partially unknown experimental environments. Performance of the proposed method is compared with the performance of the traditional particle filter approaches as the several parameters of the system are varied. These analyses show that the proposed approach improves the localization success. Additionally, the proposed method is realized by using P3-DX mobile robot platform to solve the global and position tracking localization problems. In order to provide accurate navigation a simple orientation controller is designed. The experimental results are promising for the future works.

A new approach to improve the success ratio and localization duration of a particle filter based localization for mobile robots

In real world applications, it is important that mobile robots know their location to achieve goals correctly. The localization of the robot is difficult by using raw sensor data because of the noisy measurements from these sensors. To overcome this difficulty probabilistic localization algorithm approaches can be used. The Particle filter is one of the Bayesian-based methods. In this study, two new features incorporated into the particle filter approach. These features are: decreasing the size of sample space using compass data and a new sensor model. The proposed approach is applied in the localization problem of a mobile robot. Performance of the proposed algorithm is compared with the performance of traditional particle filter approach by changing several parameters of the system. These analyses emphasized that the proposed approach improved the localization performance of the system. The results are promising for the future studies on this subject.

Constructing Topological Map from Metric Map Using Spectral Clustering

In large-scale environments, robots should have proper internal representation of the surroundings for achieving tasks such as localization, navigation, and exploration. Internal representations could be categorized in two ways: metric (grid-based) map and topological map. In this paper, we aim to generate a topological map representation (collision-free graph) of the large-scale environment from its metric map. The metric map is constructed by using laser range data with grid-line intersection algorithm. After metric map is obtained, spectral clustering is applied. Apart from the studies in literature, only the new cells obtained in a fixed time interval are employed in spectral clustering algorithm to avoid drawbacks of computational cost. Thus, the topological map grows incrementally in an online manner. In our work, we intended to represent the environment with as minimum as possible number of nodes. At the same time, the topological map must cover the entire environment. To do that, number of cluster is determined adaptively with respect to the number of new cells that can be combined in order to generate a cluster. Also, a simple heuristic is used for initialization of the k-means to avoid unrepeatable results. Lastly, obtained cluster centers are defined as nodes and they are connected to each other using minimum spanning tree algorithm. The proposed method is tested in ESOGU Electrical Engineering Laboratory building that is modelled in Gazebo simulation environment by using ROS. Experiments are conducted to demonstrate the performance of the proposed method.