A Survey on Active Simultaneous Localization and Mapping State of the Art and New Frontiers

I. INTRODUCTION
AUTONOMOUS operation in robotics applications requires robots to have access to a consistent model of thesurrounding environment, in order to support safe planningand decision making. Towards this goal, a robot must havethe ability to create a map of the environment, localize itselfon it, and control its own motion. Active SLAM refers tothe joint resolution of these three core problems in mobilerobotics, with the ultimate goal of creating the most accurate and complete model of an unknown environment. Active SLAM can be seen as a decision-making process in which therobot has to choose its own future control actions, balancingbetween exploring new areas and exploiting those already seento improve the accuracy of the resulting map model.

During the last decades, active SLAM has received increasing attention1and has been studied in different formsacross multiple communities, with the ambition of deployingautonomous agents in real-world applications (e.g., search andrescue in hazardous environments, underground or planetaryexploration). This divergence has broadened the scope of theproblem and provided a wider context, yielding numerousapproaches based on different concepts and theories that havemade the field flourish; but it also created a disconnect betweenresearch lines that could mutually benefit from each other.With this survey, we seek to fill this gap by providing a generalproblem statement and a unified review of related works.

Currently, active SLAM is at a decisive point, driven bynovel opportunities in spatial perception and artificial intelligence (AI). These include, for instance, the application ofbreakthroughs in neural networks to prediction beyond line-of-sight, reasoning over novel environment representations, orleveraging new SLAM techniques to process dynamic anddeformable scenes. Throughout this paper, we give a fresherpicture of active SLAM that goes beyond the classical —but still mainstream— entropy computation over discretizedgrids. Besides, we identify the open challenges that need tobe addressed for active SLAM to have an impact on realapplications, shaping future lines of research, and describinghow they can nourish from the cross-fertilization betweenresearch fields. Among those challenges, we emphasize theurgent need for benchmarks and reproducible research.

A. Historical Perspective

Ever since the first mobile robots were built in the late1940s, the ambition that they could perform autonomoustasks has been one of the major focuses of robotics research.To operate autonomously, a robot needs to form a modelof the surrounding environment —including localization andmapping— and perform safe navigation [1]. While the formerinvolves estimating the position of the robot and creating asymbolic representation of the environment, the latter refers to planning and controlling the movements of the robot to safelyachieve a goal location. Localization, mapping, and planninghave been often investigated in combination, resulting inmultiple research areas such as SLAM, active localization,active mapping, and active SLAM.

Localization and mapping were treated deterministicallyand solved independently until probabilistic approaches wentmainstream in the 1990s, when researchers realized that bothtasks were correlated and dependent of one another. SLAMrefers, thereby, to the problem of incrementally building themap of an environment while at the same time locating therobot within it [2]. This problem has attracted significantattention from the robotics community in the last decades;see [3]–[5] and the references therein.

SLAM, however, is a passive method and is not concernedwith guiding the navigation process. In contrast, active approaches do consider the navigation aspects of the problem.Bajcsy [6], Cowan and Kovesi [7], and Aloimonos et al. [8]were the first to study and analyze the problem of active perception (also referred to as active information acquisition [9])in the late nineties. Bajcsy [10] would later formally define itas the problem of actively acquiring data in order to achieve acertain goal, necessarily involving a decision-making process.For the cases in which the objective is to improve localization,mapping, or both, the problems are respectively referred to asactive localization, active mapping, and active SLAM.

Active mapping was the first problem to be addressed, datingback to the work of Connolly [11] in 1985. Better known sincethen as the next best view problem, active mapping tackles thesearch of the optimal movements to create the best possiblerepresentation of an environment. Subsequent examples date tothe 1990s [12]–[14], always under the assumption of perfectlyknown sensor localization. This problem has been primarilyaddressed in the computer vision community to reconstructobjects and scenes from multiple viewpoints, since the natureof the projective geometry for monocular cameras, occlusions,and limited field of view often make impossible to do it fromjust one viewpoint; see [15], [16] and the references therein.

In a similar vein, active localization aims to improve theestimation of the robot’s pose by determining how it shouldmove, assuming the map of the environment is known. Firstrelevant works can be traced back to 1998, when Fox et al. [17]and Borgi and Caglioti [18] formulated it as the problem ofdetermining the robot motion so as to minimize its futureexpected (i.e., a posteriori) uncertainty. In particular, it isin [17] where the foundations of the current workflow werelaid: (i) goal identification, (ii) utility computation, and (iii)action selection (we will extensively review these stages laterin this survey). Other relevant subsequent work can be foundin [19]–[22], but also in the related literature of perceptionaware planning [23] and planning under uncertainty [24].

Finally, active SLAM unifies the previous problems, andallows a robot to operate autonomously in an initially unknownenvironment. It refers to the application of active perception toSLAM and can be defined as the problem of controlling a robotwhich is performing SLAM in order to reduce the uncertaintyof its localization and the map representation [25]. Historically,active SLAM has been referred to with different terminology which has significantly hindered knowledge sharing and dissemination within the robotics community. Relevant seminalworks can be found under the names of active exploration [26],adaptive exploration [27], [28], integrated exploration [29],[30], autonomous SLAM [31], simultaneous planning, localization and mapping [32], belief-space planning (BSP) [33],or simply robotic exploration [34], [35]. It was not until 2002—when Davison and Murray [36] coined the term activeSLAM— that the robotics community started adopting thisnomenclature. Thrun and M ¨oller [26] demonstrate that in orderto solve robotic exploration, agents have to switch betweentwo opposite principles depending on the expected costs andgains: exploring new areas and revisiting those already seen,i.e., the so-called exploration-exploitation dilemma. The firstapproach in which a robot chooses actions that maximize theknowledge of the two variables of interest is attributed toFeder et al. [27], who also separate the procedure in threemajor stages as in [17]. Table I contains a subset of relevantworks that have followed [27]. This table differentiates themain aspects of each approach, including the type of sensors,the state representation, and the theoretical foundations.

B. About Previous Surveys

Only two works have previously addressed the problem ofsurveying active SLAM research. The first of them, publishedin 2016, is a section of a more general survey on SLAMcarried out by Cadena et al. [5]. The other, by Lluvia etal. [37], conducts a more extensive survey on “Active Mappingand Robot Exploration”. Table II summarizes the topics theyaddress, along with those covered in the present survey.

Cadena et al. [5] describe both the history and the mainaspects of the problem, and identify three open challenges:the decision of when to stop performing active SLAM, theproblem of accurately predicting the effect of future actions,and the lack of mathematical guarantees of optimality. However, the brevity of the active SLAM section prevented delvinginto a detailed discussion of the most relevant works orproviding a more unified mathematical formulation of theproblem. Moreover, since [5] was published, many relevantcontributions have been proposed and new open problemshave arisen. For instance, progress has been made on the wayuncertainties of the robot location and the map are representedand quantified. Furthermore, recent work has also opened newresearch endeavors, including deep learning (DL).

Lluvia et al. [37] also provide a thorough historical reviewand relate the different communities that have been tryingto solve this problem under different nomenclatures. Similarto [5], they do not attempt to present a unified mathematicalformulation of active SLAM nor do they cover utility computation, a field which has been mostly overlooked in the literature.They delve, nevertheless, into the optimization of vantagepoints and the trajectories to reach them, a new problem thathas attracted significant attention from the control communityand has seen many contributions in recent years. In [37], theauthors present a comparison between representative worksin active SLAM, although with a limited scope. Contrarilyto [37], we present a more complete analysis and a broaderset of open challenges, which extends the ones identified in [5].