Decentralized formation control for small-scale robot teams with anonymity

Abstract

This paper presents decentralized formation controls for a team of anonymous mobile robots performing a task through cooperation. Robot teams are required to generate and maintain various geometric patterns adapting to environmental changes in many cooperative robotics applications. In particular, all robots must continue to strive toward achieving the team’s mission even if some members fail to perform their role. Toward this end, formation control approaches are proposed under the conditions that robot teams are initially not allowed to have individual identification numbers (IDs), a predetermined leader, and agreement on coordinate systems. Therefore, all members are required first to reach agreement on their coordinate system and obtain unique IDs for role allocations in a self-organizing way. Then, employing IDs within a common coordinate system, two formation control approaches can be realized: leader-referenced and neighbor-referenced formations. Both approaches are verified using an in-house simulator and physical mobile robots. We detail and evaluate each formation control approach, whose common features include self-organization, robustness, and flexibility.

Introduction

Recently, the coordination of multiple robots has been gaining increasing attention, since robots which can perform cooperative tasks as a team offer many advantages over a single high performance robot in efficiency, low per robot cost, fault-tolerance, generality, and so on. Therefore, robot teams are expected to be deployed in a wide variety of applications including surveillance-and-security [1], object transportation [2], object manipulation [3], [4], search-and-rescue [5], [6], intelligent transportation systems (ITS) [7], [8], and exploration [9], [10]. To enable a team of multiple robots to successfully perform the assigned tasks, it is often required to generate and/or maintain geometric patterns adapting to environmental changes. Thus, this paper presents the formation control architecture and algorithm needed to coordinate multiple robot movements within a team. Specifically, formation control includes such functions as pattern generation, flocking,¹ and pattern switching. In practice, real-world applications require all robots to continue to strive toward achieving the team’s mission even if some members fail to function properly. In addition, every robot needs to move from one position to another position as quickly as possible according to the task [11]. Our goal is to develop a software framework for supporting general purpose applications of cooperative robots running the same algorithm.

Formation control of robot teams can be divided into centralized or decentralized approaches. The centralized approach relies on a specific robot to supervise the movement of the robots through a communication channel. Egerstedt and Hu [12] employed a virtual reference on the desired trajectory controlled from a remote host with which individual robots maintain their predefined positions. Belta and Kumar [13] generated smooth interpolating motion for individual robots, so that the total kinetic energy is minimized while certain constraints are satisfied. In general, a heavy computation burden is imposed on the supervising robot, which also requires tight communication with other robots. In contrast, the decentralized formation control is the coordination achieved through individual robot’s decisions.

Most research in decentralized control mainly focuses on (1) how to achieve a specific formation pattern [14], [15], [16], [17], (2) how to keep the formation pattern while flocking [1], [18], [20], [21], [22], [23], [24], [25], or (3) how to switch between formation patterns in order to adapt to an environment [26], [27]. For the first problem, Suzuki and Yamashita [14] studied the problem of generating regular polygonal shapes based on a non-oblivious algorithm with an unlimited amount of memory. To achieve the shapes, robots were required to utilize their past experience or memory. This algorithm was modified to an oblivious (or memoryless) algorithm and applied to circle formation by Defago and Konagaya [15]. Ikemoto et al. [16] proposed a biologically-inspired algorithm which enabled a robot team to form various geometric patterns. This study required robots to be initially lined up before generating a pattern. Fujibayashi et al. [17] proposed a probabilistic formation rule that controlled the number of connections between robots. However, it is generally difficult to choose the probability parameters according to the pattern and the number of robots. For the problem of flocking, two methods were implemented, the leader–follower method and the leaderless method. In the leader–follower method, a robot is selected as the moving reference point. Gervasi and Prencipe [18] proposed a computational solution based on CORDA [19] with weak assumptions such as asynchrony, anonymity, no memory and a simple behavior cycle. In their study, all followers generate a geometric pattern symmetrically with respect to the pre-selected leader. Balch and Arkin [20] studied a new paradigm of reactive behaviors for four formation patterns, where the robots were assigned roles such as leader or follower with unique IDs. Carpin and Parker [21] similarly introduced a cooperative leader following approach that could handle a heterogeneous team with different types of sensors using broadcast communication. As an extension of this approach, Parker et al. [22] introduced a tightly-coupled navigation assistance approach by a leader with rich sensing capability as the central figure of a robot team. Such strategies [21], [22] make the leader more costly and the team becomes less robust to the failure of the leader. Additional leader–follower approaches are introduced in [1], [23], [24]. An alternative approach uses no leader. Balch and Hybinette [25] proposed a physics-based flocking approach without a leader, inspired by crystal generation processes. Each robot had several local attachment sites that are attracted to other robots. Finally, for the problem of pattern switching, a graph theoretic approach was proposed by Desai [26] for switching to another geometric pattern. The approach used a control graph, which is a set of assigned targets, to define behaviors of multiple robots. Kurabayashi et al. [27] presented an adaptive transition technique to enable a team of robots to change formation by varying the phase gaps among artificial non-linear oscillators. General functionality of team organization, team maintenance, and team adaptation was addressed in [28], where Fredslund and Mataric used robots equipped with color helmets indicating their ID. When robots generate a formation, robot IDs and corresponding target points were predetermined in a particular class of formation. The leader may change according to the type of formation, and the followers must find a new neighbor in order to switch to other patterns. Lemay et al. [29] proposed a similar approach that assigned the position of the robots based on their IDs.

In contrast to most previous works, our approach begins with the following assumptions: (1) the team members do not have an external mark or ID; (2) the leader is not a priori selected; and (3) the team members are located at arbitrary distinct positions with no coordinate system agreement. Based on these assumptions, this paper presents a self-organizing team formation. Specifically, our proposed approach to formation control is divided into two strategies, the leader-referenced approach and the neighbor-referenced approach. Two potential contributions are: (1) the team is enabled to generate a variety of formations adapting to the given conditions and (2) the same or similar formations can be recovered in spite of a lack of some participating members. These features improved flexibility and robustness.

The remainder of this paper is organized as follows. Section 2 presents a self-organizing team formation definition and strategy for a small-scale team of multiple robots. In Sections 3 Leader-referenced formation control, 4 Neighbor-referenced formation control, the leader-referenced and neighbor-referenced approaches are proposed and then verified by simulations. Section 5 compares the two proposed approaches and introduces the hybrid control approach. Section 6 gives the experimental results with four physical robots based on the leader-referenced approach. Finally, the conclusion of this paper is explained in Section 7.