Chapter 2 Background and Related Work
2.3 SLAM and Navigation
2.3.3 Robot Navigation
A navigation process should be able to perform four stages: perception, localization, planning and actuation. The perception stage obtains and processes the information of sensors. Through this information, a robot can recognize objects around it, and further can recognize each object as a target or as an obstacle. Localization is a stage of determining the position and orientation of a robot with respect to its surroundings. Once the robot’s position has been determined, the path planning can be conducted. This includes finding a collision free path from a starting point to a desired point. Finally, the actuation stage is responsible for executing the plan. A sequence of actions to reach the desired position is determined in the actuation stage. As the first two stages are usually finished by the SLAM method, this section will discuss path planning and actuation.
In general, the research of navigation can be classified into two major areas: the global path planning and the local motion planning. A global path planning algorithm calculates optimal path to a desired position(Kambhampati & Davis, 1985). In global path planning, the environ- ment surrounding the robot and the position of obstacles are well known, and the robot is required to navigate to its destination by avoiding any obstacles. There are many popular global path planning methods, such as A* (Russell et al., 1995), D* (Stentz, 1994) and focused D* algorithm (Stentz, 1995). Local motion planning methods take into account unknown and changing characteristics of the environment based on the local sensory information, and dy- namically guide the robot according to the local information. Therefore, the local motion plan-
30
ning methods are more suitable and practical for autonomous robots, if the environment is dynamic and too complicated to be known.
The earliest algorithms for robot navigation were developed in a completely known environ- ment, filled with stationary obstacles whose positions were also known. Popular methods are the artificial potential field (APF) approach (Borenstein & Koren, 1989), the distance function (DF) approach (Gilbert & Johnson, 1985), vector field histogram (VFH) and VFH+ techniques (Ulrich & Borenstein, 1998), dynamic window approach (Tolman) (Fox et al., 1997), and rule based (RB) approach (W. Li, 1994).
The APF approach was introduced by Borenstein and Koren (1989), and it represents a good solution to achieve a fast and reactive response to a dynamically changing environment (Borenstein & Koren, 1991). This method acts to fill an observed environment with a potential field in which a robot is attracted to a target position and is repelled away from obstacles. At any position, a robot can calculate another position that has the global minimum repulsive force. The robot moves toward positions with minimum repulsive force and repeats this until it reaches the target. A limitation of the potential field method is the local minima problem: when attractive and repulsive forces on a robot are zero, the robot has to stop moving. As a result, the robot cannot reach the target. Since a robot’s motion in a dynamic field has a cer- tain amount of randomness due to the nature of a real circumstance, obstacle methods nor- mally cannot perform under all conditions.
Currently, control strategies occur as navigation processes. Reactive control methods are used in obstacle avoidance systems. Minguez and Montano (2000) presented a number of motion commands, which generate directions for a mobile robot to head in. A limitation of this meth-
31
od is the inability to process environmental changes. Thus, this method cannot work in a dy- namic environment. Boundary-following methods also use a common control structure (Chatterjee & Matsuno, 2001).In this method, an escape criterion is defined based on analys- ing surroundings. With this constraint, a robot will follow an obstacle boundary until the es- cape criterion is satisfied.Behaviour integration approaches were also developed. Those ap- proaches are comprised of a map that models a robot is surroundings, a planning module that outputs a correct direction or path, and a reactive module that makes the robot reactively avoiding the collision with any obstacles (Minguez & Montano, 2005; Philippsen & Siegwart, 2003). In the study of M. Wang and Liu (2008) , the local navigation problem was resolved by recording obstacle and trajectory information into a “memory grid” of the environment. Fuzzy systems were adopted, as they have the ability to treat uncertain and imprecise information using linguistic rules, also an advantage of not requiring a precise analytical model of the envi- ronment.
Noise is one factor that affects accuracy of a navigation process. When mapping an environ- ment, a mobile robot generally has to cope with different kinds of noise: noise in the odometry and sensor data. G. Dissanayake et al. (2000) proposed to remove a few landmarks to elimi- nate noise. Their results demonstrated that it is possible to remove a large percentage of landmarks without making a map building process statistically inconsistent. In their experi- ments, the SLAM method maintained efficiency by selecting of landmarks. Thrun (2001) used raw sensor data and performed a dense matching of scans. Navigational accuracies were- depent on the quality of images and the number and accuracy of points used in the resection. Although all approaches possess the ability to cope with a certain amount of noise in the sen- sor data, they assumed that the environment is almost static during the mapping process. To
32
deal with the noise and be robust to imprecision of sensory measurements, a fuzzy controller was adopted to design a behaviour based steering scheme (Jasmine Xavier & Shantha Selvakumari, 2015). Simulation results were used to demonstrate to the efficiency of a robot’s activities.
Earlier studies such as APF and VFH require quantitative formulation of the behaviour and ad- ditional computational effort in the implementation. These methods can perform well in static environment, but are not suitable for dynamic environments where there are a large amount of uncertainties. Modern behaviour based control strategies are computationally efficient. A fuzzy controller is commonly used to generate velocity and heading angles. Mostly simulation results of the controller are demonstrated, real-time experiments need to be further conduct- ed. Range sensors are mostly used in a navigation process. As visual sensors can supply in- formative data, a vision based robot controller for navigation is an increasing relevent research field.