An autonomous mobile robot is defined as a robot that has the ability to navigate in an environment and execute desired works with little or no human guidance. Previous work in intelligent robotics has proposed multistrategy learning that integrated several inference types and/or computational mechanisms into one learning system (Michalski and Tecuci, 1994). Examples consist of the integration of symbolic and numerical learning, hybrid computation of discrete and continuous spaces, integration of stochastic and deterministic heuristic searches, path planning and behavioral learning (Bianco and Cassinis, 1996), and reinforcement learning that uses value and policy iteration.

An efficient autonomous robot should be able to estimate its location during navigation and travel from one place to another in order to complete certain tasks. Also, the robot should be able to generate the map of the traversed environment corresponding to its position and posture (mapping) and calculate its position and posture based on the generated map (localization). These two processes are interdependent and usually termed as simultaneous localization and mapping (SLAM) (Tomono and Yuta, 2003; Thrun et al., 2005), which is considered to be an example of a multistrategy learning method.

Various conventional approaches for SLAM were developed, for instance, extended Kalman filter (EKF) SLAM, visual SLAM, and graph SLAM. These methods can handle uncertainty, high convergence, and do not require odometry hardware system. On the one hand, the EKF SLAM approach using maximum likelihood data associations for online SLAM and the resulted feature-based maps are used with point landmarks. On the other hand, graph SLAM overcomes the full SLAM problem in offline by obtaining all data gathered such as the current time, e.g., all poses and all features in the map during map building (Folkesson and Christensen, 2007; Kaess et al., 2008). However, all these conventional methods are computationally expensive.

Topological approaches map the environment as a set of nodes and edges (Benjamin and Yung-Tai, 1991). Nodes store information about the environment, while edges of the map contain robot’s odometry information gathered during navigation. Thus, a topological map represents the environment by only storing information of crucial places and connections between these places that will be used for navigation.

Hybrid approaches combine the topological and metric mapping methods in order to retain the advantages of both methods and compensate for the negative ones. Thrun (1998) proposed a method that uses the Occupancy Grid method in building a metric map of the environments, and a topological map is a posterior calculated from the metric map. However, the accuracy and consistency of the map generated by this approach are affected by accumulated odometric error present in the metric map. In Taylor and Kriegman (1998), the environment is represented as a collection of local metric maps, all of which are stored as nodes in a global topological map. The node in the topological map of this approach stores a set of landmarks observed from the boundary of each obstacle detected by the robot and uses this information for navigation between nodes in the map in the later stage. This method neglects robot pose estimation entirely.

Recently, vision-based navigation system was introduced for hybrid map building (Cristforis et al., 2015). However, this approach is based on teach-and-replay technique and requires odometry system for robot localization. Previous works such as Battesti et al. (2011) and Bazeille and Filliat (2011) used vision-based hybrid map building to create consistent topometric map and compensate odometry drift. However, these methods require robot odometry system for topometric map building.

In this article, we proposed a new approach that integrated our previous work SSGA (Toda et al., 2012) and Bayesian ARAM with map maintenance for hybrid map building. The Bayesian ARAM with map maintenance is the extension of our previous work (Chin et al., 2015) that is crucial for dealing with proliferation problem of nodes. The proposed method termed as Genetic Bayesian ARAM, which constructs map, contains both a metric and a topological representation. First, SSGA allows the robot to perform self-localization and metric mapping simultaneously based on occupancy grid mapping. Then, Bayesian ARAM utilizes the localization information from the metric map and sensory information from the explored environment to construct the topological map. Topological nodes represent distinct places, while edges connect nodes and store robot’s bearing information such as orientation and direction. The metric grid map describes the explored environment outline for understanding of human operators and further operations. In addition, it also provides global position of robot for Bayesian ARAM to generate the topological map and overcome the online detection and recognition problem. With the topological map, the robot could perform localization and path planning without recalculating the entire metric grid map. In addition, previous works such as Duan et al. (2015) and Wu et al. (2013) have used genetic algorithm for SLAM purpose. Table 1 shows the difference between the previous two works and our proposed method.

Contributions of this paper are: (i) it does not require odometry system such as encoder and GPS for robot position estimation; (ii) the unsupervised online learning feature enables the robot to build the topological metric map from scratch and maintain it naturally without any human interference; and (iii) it does not require any artificial marker or adjustment and complicated cognitive intelligence for operating in a natural environment. Section 2 introduces the theoretical framework of the proposed simultaneous localization and hybrid map building approach. The experimental setup is explained in Section 3, and experimental results are discussed in Sections 4 and 5. Finally, conclusions are presented in Section 6.

Our proposed method termed as Genetic Bayesian ARAM is the integration of our previous works (Toda et al., 2012; Chin et al., 2015). In this paper, we have further improved our proposed method with map maintenance feature, which overcomes the proliferation problem of nodes. In our proposed hybrid mapping method, the notion of map is described by both a metric and a topological feature. On the metric side, the environment is represented by a global occupancy grid map. Steady-state genetic algorithm (SSGA) simultaneously performs self-localization and metric map building based on occupancy grid mapping.

On the other hand, Bayesian ARAM obtains robot’s location from the metric mapping and sensory information to build the topological map. Bayesian ARAM continuously clusters sensory sources as nodes and creates edges to connect one another if the robot traverses between these nodes. Each of these connections contains robot’s movement information, while nodes represent important places of the explored surroundings. The proposed method continuously updates the hybrid map representations with little or no human intervention. The overall process of our proposed framework is shown in Figure 1.

This section shows the setup of experiment, while Section 4 shows the results collected from the simulation and real robot implementation. The hybrid map is formed by the occupancy grid map and topological map. For the occupancy grid map, white color represents empty space, gray color represents unknown space, and black color represents obstacles or boundary detected, whereas the topological map represents set of nodes and links. Red color circles at the (x, y) coordinates and black color lines represent nodes and links in the map, respectively.

According to Vigdor and Lerner (2007) and Chin et al. (2016), we have mentioned the optimized range of parameters’ setting. Therefore, we follow the parameters’ setup mentioned in paper (Vigdor and Lerner, 2007; Chin et al., 2016) because the authors have determined the optimized value for the Bayesian ART parameters; therefore, we will not cover in details in this paper.

First, we set the Thales sphere radius e_max as 0.3 and the initial covariance matrix to be spherical Σinit=σinit2I, where I is the identity matrix; hence, the parameter σinit2≪(SMAX)1∕M [equations (13) and (14)] to ensure that the σinit2 is always lower than the S_MAX. Then, we set the maximal category hypervolume parameter value (S_MAX) as 1. Next, the value of the initial parameter (σinit2≪11∕682) was set to 0.01. Note that 682 is the laser scanner scans. Finally, we set the P^(wj)init to 1 for generating the deterministic region instead of probabilistic region map.

As mentioned in the previous section, we have configured Genetic Bayesian ARAM to two channels for attaining laser rangefinder data and robot location in the experiment. Thus, the importance factor α for laser rangefinder channel was set to 0.8 because its data are the main dominant for map building. Last, the importance factor α was set to 0.2 for robot location channel to deal with perceptual ambiguities of sensory information but not the environment condition. For SSGA variables’ setup, we set parent candidates (μ) = 1000 and offspring candidates (λ) = 500.

In Section 4.1, the performance of Genetic Bayesian ARAM in a hybrid map building was validated using several benchmark datasets. The benchmark datasets were produced by the University of Freiburg, Department of Computer Science, with the objective for contributing benchmarking tools to robotics community.

For the physical robot implementation, the front laser rangefinder data, rear laser rangefinder data, and robot odometry system that was installed on the robot are transmitted to Genetic Bayesian ARAM for map building and learning. Section 4.3 explains details of the setup.

We have shown that the Genetic Bayesian ARAM framework is capable of generating a metric–topological hybrid map incrementally using nearly unprocessed sensory information and without any feature extraction and predefined knowledge about the environment. The algorithm updates both the metric map and creates nodes incrementally to represent the explored environment as regions in real time.

In all the experiments, SSGA continuously performed localization and metric mapping without using any odometry system. Then, Bayesian ARAM uses these localization data and laser scanner data for topological map building. The localization data are important for distinguishing places that have similar sensor data, which is important to overcome the online topological nodes’ localization. After building the map, the robot can perform path planning by just comparing its current location with topological nodes without recalculating the grid map. These features help to compensate the weakness of metric map and topological map.

In addition, the experiment showed that the map maintenance algorithm greatly reduces the number of nodes that represent the explored environment. This method also reduces the amount of memory required for storing the map, especially when the area is very large.

According to the node matching and localization results, the robot failed for localization occasionally during navigation because it encountered sudden change in the environment, which caused none of topological nodes match with these sensor data. This problem can be overcome by measuring few time stamps of sensor at the same position with the topological nodes in the map. Another reason is that the robot was unable to differentiate similar sensor data at particular places. As our proposed method can take more than one sensory source, this issue can be solved by integrating camera input with certain image processing techniques to help on disambiguate similar sensory information at different locations.

In this article, we proposed a new framework for simultaneous localization and hybrid map building without any odometry hardware. The experiments presented show the feasibility of the proposed approach. The proposed framework was verified through simulation of standardized benchmark datasets and real robot application. Both metric and topological maps were generated as expected, correctly representing the environment according to our proposed method. The algorithm does not require high-level cognitive knowledge or any artificial landmark to construct the hybrid map, which makes it ready to work in a natural environment. In addition, the robot can perform self-localization automatically without any odometry system.

Future work in this subject will include an analysis of effectiveness of the value of framework parameters. Besides, we will extend our method for reinforcement learning capability for path planning. Last, we will conduct more experiments in different kinds of indoor and outdoor environment for further verification.

WC is the main contributor for this research and paper writing. YT is the contributor for developing steady-state genetic algorithm and grid map building. Dr. CL and Dr. NK are the supervisors.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.