Evolutionary Robotics 1

Evolution is a widely used paradigm of computing that is finding an application in all domains. Evolutionary algorithms use an analogy from natural evolution to model the problem solving approach. They consist of a set of individuals that are the solutions to the problem and make a population pool. The various operators are used to modify the population pool of individuals. This consists of scaling, selection, mutation, crossover, and elite. Evolutionary algorithms are used for the problem of motion planning of the robots. Non-holonomicity is a major issue associated with mobile robots. The paths returned by the planners need to be smooth to ensure easy tracking by the robot control algorithms. The authors make use of BSplines and Bezzier curves to solve the problem. These are smooth curves that are controlled by the control points on the maps. Adaptation is another major aspect studied in evolutionary algorithms. The authors use evolutionary strategies to solve the problem of motion planning. These are self-adaptive in nature and are able to adapt the evolutionary parameters during optimization. Covariance Matrix Adaptation Evolutionary Strategy is the studied method of implementation.

Conclusion

Having covered a large number of concepts, algorithms, issues, and challenges in robot motion planning, there is little left to state as the conclusion. In this chapter, the authors simply jot down some last words, final thoughts, and closing remarks. While a large number of algorithms exist for the same problem, it is hard to pick out one algorithm and use it upright for a given modeling scenario. It becomes important to understand what features the algorithm has, what is the level of each of these features, and whether these features suit the problem in hand. Optimality and completeness are important factors provided by the algorithm in real time, which may, however, not always be possible. Hybrid algorithms are more tailored in the sense that these algorithms can be customized as per needs. Again, the choice of algorithm plays an important role, as the limitations get added up along with the advantages which we need by design. Hence, it is always important to know what is desired from the robot and the planning algorithm and what scenarios it may face. Most practical scenarios may be simple, such that these can be simply solved by any general algorithm. Other aspects of robotics may be more challenging than a planning algorithm. However, it is certain the ever-rising use of robotics shall bring forth challenging scenarios for robots to work in, in which case the planning algorithms would be tested to best of their capability.

Multi-Robot Systems

While the concepts of robotics and planning may be easily understood by the taking a single robot, it is not necessary that the problems we solve have a single robot in the planning scenario. In this chapter, the authors present systems with multiple robots, each robot attempts to coordinate and cooperate with the other robots for problem solving. The authors first look at the specific problems where multiple robots would be a boon for the system. This includes problems of maze solving, complete coverage, map building, and pursuit evasion. The inclusion of multiple robots in the scenario takes all the concepts of single robotic systems. It also introduces some new concepts and issues as well. They look into all these issues in the chapter which include optimality in terms of computational time and solution generated, completeness of planning, reaching a consensus, cooperation amongst multiple robots, and means of communication between robots for effective cooperation. These issues are highlighted by specific problems. The problems include multi-robot task allocation, robotic swarms, formation control with multiple robots, RoboCup, multi-robot path planning, and multi-robot area coverage and mapping. The authors specifically take the problem of multi-robot path planning, which is broadly classified under centralized and decentralized approaches. They discuss means by which algorithms for single robot path planning may be extended to the use of multiple robots. This is specifically done for the graph search, evolutionary, and behavioral approaches discussed in the earlier chapters of the book.

Hybrid Evolutionary Methods

The limitations of single algorithm approaches lead to an attempt to hybridize or fuse multiple algorithms in the hope of removing the underlying limitations. In this chapter, the authors study the evolutionary algorithms for problem solving and try to use them in a unique manner so as to get a better performance. In the first approach, they use an evolutionary algorithm for solving the problem of motion planning in a static environment. An additional factor called momentum is introduced that controls the granularity with which a robotic path is traversed to compute its fitness. By varying the momentum, the map may be treated finer or coarser. The path evolves along the generations, with each generation adding to the maximum possible complexity of the path. Along with complexity (number of turns), the authors optimize the total path length as well as the minimum distance from the obstacle in the robotic path. The requirement of evolutionary parameter individuals as well as the maximum complexity is less at the start and more at the later stages of the algorithm. Momentum is made to decrease as the algorithm proceeds. This makes the exploration vague at the start and detailed at the later stages. As an extension to the same work, in the second approach of the chapter, the authors show the manner in which a hybrid algorithm may be used in place of simple genetic algorithm for solving the problem with momentum. A Hybrid Genetic Algorithm Particle Swarm Optimization (HGAPSO) algorithm, which is a hybrid of a genetic algorithm and particle swarm optimization, is used in the same modeling scenario. In the third and last approach, the authors present a hierarchical evolutionary algorithm that operates in two hierarchies. The coarser hierarchy finds the path in a static environment consisting of the entire robotic map. The resolution of the map is reduced for computational speed. The finer hierarchy takes a section of the map and computes the path for both static and dynamic environments. Both these hierarchies carry optimization as the robot travels in the map. The static environment path gets more and more optimized along with generations. Hence, an extra setup cost is not required like other evolutionary approaches. The finer hierarchy makes the robot easily escape from the moving obstacle, almost following the path shown by the coarser hierarchy. This hierarchy extrapolates the movements of the various objects by assuming them to be moving with same speed and direction.

Behavioral Path Planning

A large part of our everyday motion is governed by behaviors. We necessarily do not look at the entire map and formulate the best way out, but rather take instinctive actions regarding our motion. We naturally reach the desired locations fairly easily and near optimally. With the same inspirations in mind, in this chapter, the authors explore the behavioral systems for the task of motion of the mobile robot. In this chapter, they study two different algorithms, fuzzy inference systems and artificial neural networks. The fuzzy systems are governed by a set of rules, which determine the behavior of the system, for any applied input. The major task involves the use of fuzzy sets for the output computations. As per the theory of these sets, every input belongs to every set by a varying degree called as the membership degree. The authors use this concept of fuzzy-based inference to design a system for the motion of the mobile robot. They further introduce the neural networks paradigm, which is an inspiration from the human brain for problem solving. Neural networks process applied input layer wise by unit processing centers known as artificial neurons. These systems may be trained by a training database that is particular to a problem. The authors use both these algorithms to design systems for behavioral path planning of mobile robots.

Evolutionary Robotics 2

Evolution is a major computing tool. Its effectiveness can be seen in a large variety of problems. A number of evolutionary algorithms exist that differ in the mechanism in which they represent the problem and carry out the evolutionary process. In this chapter, the authors extend the ideas of the previous chapter to look for more evolutionary concepts and algorithms. The first major class includes swarm intelligence algorithms, where they primarily study particle swarm optimization, ant colony optimization, artificial bee colonies, and probability-based incremental learning. In all these algorithms, the basic motive is to use a number of particles that form a complete population, and optimization is carried out by their mutual interaction. The other major algorithm that the authors study in the chapter is Genetic Programming, which forms a major pillar of evolutionary computation. Here, the individual is a program whose execution results in the estimation of fitness denoting the goodness of problem solving. The various operators are modified to enable the representation of the individual to work. A tree-based representation is one of the most commonly used representations. They further study a linear representation of the program, and the underlying technique is known as grammatical evolution. All the algorithms are applied to the problem of motion planning of mobile robots.

Introduction

The basic purpose of the book is to specifically look into the problem of robot planning, which is an essential problem in robotics. Before the authors discuss the simple and complex algorithms, analyze the same, present the various modeling scenarios, and present some results and associated limitations, tradeoffs, and issues, it is wise to first understand the basics of robotics. The adoption of Artificial Intelligence in robotics is a never-ending effort, a part of which is the crux of the book. It is important to understand the notion of problem of path planning and where and how it fits into the entire domain of robotics. The chapter is an introduction to the entire book and gives an overview of all the background needed, including the underlying problem of motion planning of mobile robotics. In other words, this chapter gives a broader view of the complete system, a part of which, will be looked at in intensive details in this book.

Common Planning Techniques

Motion planning of mobile robots can be done by a variety of mechanisms, which makes it a very interesting field of work. In this chapter, the authors discuss some of the commonly used techniques that resemble or closely follow the graph-based search techniques discussed in chapter 2. They take a detailed study of the use of algorithms, like dynamic programming, Bellman Ford algorithm, Rapidly exploring Random Trees (RRT), artificial potential fields, embedded sensor planning, reinforcement learning planning, and Voronoi graph-based planning. Bellman Ford is a graph search algorithm that finds the shortest path between the source and all the vertices in graph. RRTs are tree-like data structures where every node is a free position in the map. The trees grow to explore areas and thereby find the path from source to goal. The artificial potential field method uses the concept of potential by every obstacle and goal for motion planning of the robot. The planning in embedded sensor networks is distributed in the sensors distributed in the entire map. The discussions enable a multi-dimensional outlook over the problem. The authors note the similarities, differences, advantages, and the disadvantages between the different algorithms.

Hybrid Behavioral Methods

The chapter focuses upon the use of hybrid planning systems that are primarily behavioral in nature. Three approaches are discussed. In the first approach, the authors solve the problem of path planning using a combination of A* algorithm and Fuzzy Inference System. The A* algorithm does the higher-level planning by working on a lower detail map. The algorithm finds the shortest path, while at the same time generating the result in a finite time. The A* algorithm is used on a probability-based map. The lower-level planning is done by the Fuzzy Inference System (FIS). The FIS works on the detailed graph where the occurrence of obstacles is precisely known. The FIS generates paths that take into account the non-holonomic constraints and generate smoother paths. The results of A* algorithm serve as a guide for FIS planner. The FIS system is initially generated using rules from common sense. Once this model is ready, the fuzzy parameters are optimized by Genetic Algorithm (GA). The GA tries to optimize the distance from the closest obstacle, total path length, and the sharpest turn at any time in the journey of the robot. Many times, the planning algorithm may require a map breakup such that the coarser-level graph also has a high degree of resolution and A* algorithm may not be able to work. Hence, in the second approach, the authors replace the A* algorithm with GA. Both approaches were tested on various complex and simple paths. All paths generated were optimal in terms of path length and smoothness. Their last approach is based on dynamic programming, which, in this implementation, is similar to the use of neurons embedded in the map being planned. In this approach, the authors talk about the use of extra nodes in the planning framework called accelerating nodes. These nodes are less in number and fully interconnected. These nodes transmit information about any map change and blockages to each other for sudden re-planning to be initiated. These nodes further guide the robot until re-planning completes.

Graph Based Path Planning

Graphs are keenly studied by people of numerous domains as most of the applications we encounter in our daily lives can be easily given a graph-based representation. All the problems may then be easily studied as grap-based problems. In this chapter, the authors study the problem of robot motion planning as a graph search problem. The key steps involve the representation of the problem as a graph and solving the problem as a standard graph search problem. A number of graph search algorithms exist, each having its own advantages and disadvantages. In this chapter, the authors explain the concept, working methodology, and issues associated with some of these algorithms. The key algorithms under discussion include Breadth First Search, Depth First Search, A* Algorithm, Multi Neuron Heuristic Search, Dijkstra’s Algorithm, D* Algorithm, etc. Experimental results of some of these algorithms are also discussed. The chapter further presents the advantages and disadvantages of graph-based motion planning.