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

I. INTRODUCTION

• Active SLAM jointly resolves mapping, localization, and motion control, while choosing actions that balance exploring new areas with improving maps of areas already seen.Its goal is to create an accurate and complete model of an unknown environment.Source: §I · p. 1, FSE T45 · p. 1

FSE T45

• The survey aims to bridge research lines by providing a general problem statement and unified review of approaches developed across multiple communities.The field’s divergence has broadened its scope but also created a disconnect between lines of work that could benefit from one another.Source: FSE T45 · p. 1
• The survey examines new opportunities in spatial perception and AI, identifies challenges for real-world impact, and emphasizes the urgent need for benchmarks and reproducible research.Topics include neural-network prediction beyond line of sight, new environment representations, and SLAM techniques for dynamic and deformable scenes; the review goes beyond entropy computation over discretized grids.Source: FSE T45 · p. 1

A. Historical Perspective

• Autonomous robots need an environment model for localization and mapping, as well as planning and control to navigate safely toward a goal.Localization estimates robot position and creates a symbolic environment representation; planning and control guide movement toward the goal.Source: A. Historical Perspective · pp. 1–2
• Publications on active SLAM grew twelve-fold, from 53 in 2010 to over 660 in 2022, and reached almost 5500 when related planning and exploration terms were included.Exploration requires balancing new-area exploration against revisiting known areas; Feder et al. proposed actions maximizing knowledge of both variables of interest and separated the process into three stages.Source: A. Historical Perspective · pp. 1–2
• In the 1990s, probabilistic approaches replaced independent deterministic treatment by recognizing localization and mapping as correlated; SLAM incrementally builds a map while locating the robot.SLAM drew sustained attention from robotics researchers –.Source: A. Historical Perspective · p. 2
• Active perception adds navigation decisions to passive SLAM; early studies appeared in the late 1990s, and active perception was defined as goal-directed data acquisition involving decision-making,,,,.Depending on the objective, these problems are called active localization, active mapping, or active SLAM.Source: A. Historical Perspective · p. 2
• Active mapping dates to Connolly’s 1985 work and seeks movements that produce the best environment representation, assuming known sensor localization in early examples –.It was primarily studied in computer vision, where occlusions and limited monocular-camera field of view make reconstruction from one viewpoint difficult.Source: A. Historical Perspective · p. 2
• In 1998, Fox et al. and Borgi and Caglioti framed active localization as choosing motion to minimize future expected uncertainty; established goal identification, utility computation, and action selection.Later work included perception-aware planning and planning under uncertainty, alongside studies –.Source: A. Historical Perspective · p. 2
• Active SLAM unifies active mapping and localization to reduce uncertainty in robot pose and map representation while operating in an initially unknown environment.The term was coined by Davison and Murray in 2002, after varied labels—including active exploration, adaptive exploration,, integrated exploration,, autonomous SLAM, simultaneous planning, localization and mapping, belief-space planning, and robotic exploration, —hindered knowledge sharing.Source: A. Historical Perspective · p. 2

B. About Previous Surveys

• Only two earlier works surveyed active SLAM: Cadena et al. in a general SLAM survey and Lluvia et al. on active mapping and robot exploration.Table II compares their topics with those covered by the present survey.Source: B. About Previous Surveys · p. 2
• Cadena et al. reviewed active SLAM history and challenges, but the section’s brevity precluded detailed work coverage or a unified mathematical formulation.Since its publication, research has advanced uncertainty representation and quantification and opened new directions including deep learning.Source: B. About Previous Surveys · p. 2
• Lluvia et al. reviewed history and vantage-point and trajectory optimization, but omitted unified formulation and utility computation and offered a limited comparison.The present survey provides a more complete analysis and broader set of open challenges than, extending those identified in.Source: B. About Previous Surveys · p. 2

& IMU

• The supplied introduction identifies a historical review, and Table I is described as a chronological comparison of representative active SLAM approaches.Source: Table I · p. 3, & IMU · p. 4

Problem formulation

• The survey flags novel representations, reproducible research, and practical applications, and compares topics and open challenges covered by earlier surveys and the current work.Source: Problem formulation · p. 4, Table II

C. Paper Structure

• The paper’s structure assigns Sections III–V to action identification, utility computation, and action selection and execution, then covers alternative methods, multi-robot work, and open questions.Sections VI–VII address continuous-state optimization and deep-learning methods; Section VIII covers multi-robot active SLAM, and Section IX outlines open questions.Source: C. Paper Structure · p. 4

A. Problem Formulation

• Active SLAM is framed as a POMDP whose uncertain state is represented by a belief, with the SLAM state combining robot state and map.The belief is a sufficient statistic; active-SLAM rewards typically reflect belief uncertainty rather than reaching specific states.Source: A. Problem Formulation · pp. 4–5

B. Decoupling Active SLAM into Three Subproblems

• The traditional modular approach identifies candidate actions, estimates their costs and gains, then selects and executes actions iteratively until stopping conditions are met.Decoupling can produce suboptimal results; continuous-action optimization and POMDP-based policy optimization or learning can instead handle the stages together.Source: B. Decoupling Active SLAM into Three… · p. 5

III. STAGE 1: IDENTIFICATION OF POTENTIAL ACTIONS

• Candidate-action identification reduces and discretizes the search space; frontiers became the most common approach after earlier use of random or human-selected goals.The section introduces goal-identification methods and map representations on which they depend.Source: §III · p. 5

A. Representation of the Environment

• The survey reviews four environment-map representation types: topological, metric, metric-semantic, and hybrid maps.Source: A. Representation of the Environment · p. 5

1) Topological maps

• Topological maps represent convex free-space regions as graph vertices and their connections as edges, typically segmenting an occupancy grid; despite graph-planning benefits, they are uncommon in active SLAM.The passage points to for a survey of graph-construction methods and cites, on their use in active SLAM.Source: 1) Topological maps · p. 5

Metric maps

• Sparse maps represent scenes with interest points or landmarks and are used in optimal control,, and belief-space planning.Source: 2) Metric maps · p. 5
• Dense maps include point clouds, meshes, and cell-based metrics; occupancy grids encode each cell’s probability of being occupied and extend to 3D forms such as OctoMaps, Supereight, and voxel maps.Occupancy grids have been used in numerous active SLAM frameworks,,,, as have 3D extensions –; continuous occupancy maps support continuous optimization.Source: 2) Metric maps · p. 5
• Signed-distance-field maps such as Voxblox encode more sophisticated metrics but are seldom used in active SLAM.Source: 2) Metric maps · p. 6

3) Metric-semantic maps

• Semantic maps represent objects by category, pose, and shape, but active SLAM use remains limited apart from informative-trajectory selection with a semantic OctoMap,.Dense metric-semantic maps include SDF-based Voxblox++ [95] and Kimera, and voxel-based Fusion++ [97] and.Source: 3) Metric-semantic maps · p. 6

4) Hybrid and hierarchical maps

• Hybrid and hierarchical maps combine metric, semantic, and topological information, but their use in active SLAM remains mostly unexplored, with among the few examples.Rosinol et al. combine these representations in a 3D scene graph for high-level reasoning.Source: 4) Hybrid and hierarchical maps · p. 6

B. Detecting Goal Locations

• Because exhaustively identifying destinations is intractable, methods use finite candidate sets; frontier exploration is common, while random selection assumes equal information at every location.Finite sampling improves tractability but does not guarantee global optimality; frontiers mark boundaries between known and unknown regions and were introduced by Yamauchi.Source: B. Detecting Goal Locations · p. 6
• For 2D maps, frontier detectors reduce search through free-space, recent-scan, or incremental strategies; 3D methods often evaluate map portions incrementally or along surfaces.Frontier clustering, such as K-means or mean-shift, can reduce the dimensionality of detected sets.Source: B. Detecting Goal Locations · p. 6
• Frontier detection can incorporate loop-closure regions because, under high uncertainty, revisiting or overlapping sensor views may aid localization and mapping more than exploring new space.Active SLAM approaches include loop-closure candidates alongside frontiers, or switch between frontier exploration and revisiting known places,,.Source: B. Detecting Goal Locations · p. 6
• Local goal selection can yield only locally optimal decisions and short horizons may cause poor behavior, while RRT-based candidate paths and continuous-domain optimization consider broader spaces.This local strategy is typical of deep reinforcement learning, where network memorization can mitigate local optimality –.Source: B. Detecting Goal Locations · p. 6

IV. STAGE 2: UTILITY COMPUTATION

• Utility computation evaluates each destination’s expected effect and typically aggregates travel cost with mapping and localization uncertainty, though some methods instead seek Pareto-optimal multi-objective solutions –.The appropriate metric depends on the variable representation: information-theoretic metrics usually target OG maps, while optimal-design metrics suit Gaussian distributions.Source: §IV · pp. 6–7

A. Naive Cost Functions

• Naive costs use distance, travel time, or expected area visited; they are representation-agnostic, but combining distance with information-based terms commonly requires manual tuning,.Expected visited area approximates map entropy in OG maps, and combining distance and information has been shown to improve exploration,.Source: A. Naive Cost Functions · p. 7

B. Information Theory (IT)

• Entropy, introduced by Shannon, measures uncertainty and relates to information; early map-only exploration,, ignored localization error, distorting expected map uncertainty.Source: B. Information Theory (IT) · p. 7
• Because joint entropy is generally intractable, graph-SLAM methods approximate pose and map entropies separately, while particle-filter SLAM uses a weighted mean over particles.The separate terms can have mismatched magnitudes, addressed with weights,, combined Shannon–Rényi utility,, or EM virtual maps [136].Source: B. Information Theory (IT) · p. 7
• Robot-state entropy can use the posterior covariance log-determinant under a Gaussian assumption, while map entropy depends on representation, including independent-cell occupancy entropy for grid maps.Landmark-map entropy can use the same Gaussian assumption; occupancy entropy has been used for 2D and 3D occupancy grids, with nearby-cell evaluation proposed for particle-filter SLAM.Source: B. Information Theory (IT) · p. 7
• Mutual information (MI) is the most common active-SLAM utility, measuring expected entropy reduction after an action over possible future measurements.The expectation is taken with respect to expected future measurements.Source: B. Information Theory (IT) · p. 7
• Kullback–Leibler divergence (KLD) measures distributional change in both form and mean, comparing prior and posterior distributions or estimated and true posteriors.Source: B. Information Theory (IT) · p. 7
• For occupancy-grid maps, entropy, MI, and KLD rely on cell counts and are discrete, making them ill-suited to optimization; a differentiable alternative still assumes perfect localization.The proposed cost-utility function applies to both 2D occupancy grids and voxel maps,.Source: B. Information Theory (IT) · p. 8
• Focused active inference reduces uncertainty in a selected variable subset, which may differ from the subset favored by maximizing information over all variables.It is generally more computationally intensive because it requires marginalizing the posterior Fisher information matrix; a matrix-determinant-lemma method avoids computing posterior covariance for focused and unfocused entropy,.Source: B. Information Theory (IT) · p. 8

C. Theory of Optimal Experimental Design (TOED)

• TOED compares actions by posterior covariance, with T-, D-, A-, and E-optimality capturing different variance or covariance-ellipsoid properties; monotonicity depends on uncertainty representation.Differential representations guarantee monotonicity for all utility functions, while some other representations can lead to incorrect exploration decisions.Source: C. Theory of Optimal Experimental… · p. 8

D. The Graphical Structure of the Problem

• Graph connectivity can stand in for optimality criteria under stated pose-graph conditions, linking D- and E-optimality to spanning trees and algebraic connectivity.Extensions connect T-optimality to average node degree and report substantial computation-time reductions for appropriately weighted SE(n) pose graphs.Source: D. The Graphical Structure of the… · pp. 8–9

V. STAGE 3: ACTION SELECTION AND EXECUTION

• Discrete destination selection can enumerate candidate goals, but decoupling goal choice from path planning means the executed path is rarely optimal for the original problem.After selecting a destination, methods such as RRT and PRM can plan a path to it.Source: §V · p. 9

VI. BELIEF-SPACE PLANNING AND CONTINUOUS-SPACE OPTIMIZATION; A. Belief-Space Planning (BSP)

• Belief-space planning optimizes future robot trajectories either over candidate paths or continuously, while exact continuous-state POMDP solutions are generally intractable.Candidate-path methods evaluate utility along paths, and continuous BSP methods optimize without discretizing actions; BRM-based approaches account for predicted position or pose uncertainty.Source: §VI · p. 9, A. Belief-Space Planning (BSP) · pp. 9–10

B. Active SLAM as Optimal Control

• With linear transition and observation models, Gaussian noise, and covariance-dependent reward, active SLAM reduces to deterministic optimal control; constraints and loop closures remain challenges.When deterministic reduction assumptions fail, an open-loop sequence can be followed by closed-loop feedback; constrained formulations include DDP, MPC, and augmented-Lagrangian methods.Source: B. Active SLAM as Optimal Control · p. 10

VII. DEEP-LEARNING-BASED APPROACHES; A. Deep Reinforcement Learning (DRL)

• Deep-learning approaches embed utility computation and action selection in a network, with DRL prominent because agents must learn from interaction, partial observations, and scenarios lacking prior knowledge.DRL testing can require only a network forward pass, but behavior depends on training and may not generalize to novel operational conditions.Source: §VII · p. 10, A. Deep Reinforcement Learning (DRL) · p. 10

B. On the Reward Function Design and the Action Set

• Extrinsic-reward exploration can address obstacle avoidance rather than active SLAM, prompting intrinsic coverage, curiosity, and uncertainty rewards, while rewards capturing both localization and mapping uncertainty remain open.Some methods use learning to select among frontiers or sampled locations while leaving motion planning outside learning; others combine long- and short-term policies with a model-based planner.Source: B. On the Reward Function Design and… · pp. 10–11

C. Partial Observability and Generalization

• Partial observability makes learned policies suboptimal from single observations, while generalization to unseen environments remains a key limitation and transfer to real scenarios an open problem.History inputs and recurrent networks can address partial observability; random starts, observation noise, sparse or egocentric inputs, and graph representations can mitigate overfitting, but few efforts transfer trained agents to real scenarios.Source: C. Partial Observability and… · p. 11

D. Training Environments

• DRL training needs online simulation because offline-data training is unavailable and real-world training appears infeasible; simplified scenarios can restrict inputs to ground truth or perfect range observations.More complex simulators are needed to provide realistic data that bridge the gap between simulation and physical robots.Source: D. Training Environments · p. 11
• Robotics simulators trade off capability and integration: Stage is 2D, Gazebo supports 3D and ROS, and CoppeliaSim/V-REP offers mesh manipulation but is less ROS-integrated and not open source.Zamora et al. combine OpenAI Gym with ROS and Gazebo to address integration between physics engines and DRL frameworks.Source: D. Training Environments · p. 11
• AI-era simulators prioritize training speed; DeepMind Lab, Habitat-Sim, and iGibson offer different sensing and simulation capabilities, but none has been used for active-SLAM DRL.Habitat-Sim can render image-dataset environments such as Replica, while iGibson includes lidar and optical flow simulation and ROS integration; Habitat-Sim requires external libraries for ROS.Source: D. Training Environments · p. 11

VIII. MULTI-ROBOT ACTIVE SLAM

• Joint multi-robot active SLAM scales exponentially with team size, making centralized planning unsuitable for large, intermittently connected teams and motivating decentralized or distributed methods.Centralized planning can suit small teams with continuous connectivity; Kantaros et al., propose informative planning that scales to many sensors and targets and is probabilistically complete and asymptotically optimal.Source: §VIII · p. 12
• Collaborative exploration coordinates frontier utility and travel cost through decentralized objectives, trajectory-space planning, anytime search, RRT*, ρ-POMDPs, and hierarchical region assignment with DRL navigation.Examples include entropy-distance objectives that penalize regions other robots explore, coordinate descent over agent trajectories, real-time search, exploration-biased RRT* [225], cooperative planning with periodic communication, and region assignment with collision-free DRL navigation.Source: §VIII · p. 12
• Collaborative multi-robot active estimation selects actions to reduce uncertainty in localization, maps, future observations, or other robots’ pose graphs.Methods include relative-entropy trajectory selection, collaborative belief-space planning, a decentralized extension, self-organizing maps for active perception, and graph-connectivity-based pose-graph improvement through shared observations.Source: §VIII · p. 12

IX. OPEN RESEARCH QUESTIONS

• Active SLAM still needs research to support fully autonomous robots in complex environments, and the section frames selected research questions as requiring cross-field cooperation for real-world impact.The authors note that some challenges are long-known and intensely investigated, while others have received less attention.Source: §IX · p. 12

A. Prediction Beyond Line-of-sight

• Deep-learning methods predict beyond-line-of-sight maps or scenes using augmented depth, occupancy grids, variational auto-encoders, and 3D SDF maps,,,,,.Traditional prediction uses sensor models with ray casting.Source: A. Prediction Beyond Line-of-sight · pp. 12–13
• These approaches are promising for online prediction but remain unintegrated into active SLAM; open issues include uncertainty, unstructured scenes, measurement-versus-map prediction, and action-conditioned forecasts,,.The section also asks whether predictions should augment the whole scene or reflect only selected non-myopic actions.Source: A. Prediction Beyond Line-of-sight · p. 13
• Robot motion is readily predicted, but trajectory-uncertainty forecasting remains difficult; Asraf and Indelman use data-driven scene prediction with BSP and predicted observations, while loop-closure effects remain open,.Loop closures can significantly affect uncertainty in the robot’s new states.Source: A. Prediction Beyond Line-of-sight · p. 13

B. From Active SLAM to Active Spatial Perception

• Modern SLAM builds metric-semantic maps and hierarchical scene graphs,,, but using these representations for active SLAM remains largely unexplored.Semantics may improve unseen-space prediction and SLAM, while hierarchy may enable more efficient planning; uncertainty, computation, and spatial prediction remain open questions.Source: B. From Active SLAM to Active Spatial… · p. 13

C. Robust Online Belief Space Planning and Active SLAM

• Ambiguous data association is only recently being modeled in active SLAM and belief-space planning, but enumerating hypotheses grows exponentially and makes optimal action planning intractable.Pathak et al. reason about future association hypotheses; Shienman and Indelman use a distilled subset with guarantees in a myopic setting, while broader problem simplification remains open.Source: C. Robust Online Belief Space Planning… · p. 13

D. Reasoning in Dynamic and Deformable Scenes

• Active SLAM for dynamic scenes has largely focused on replanning during execution,, while no framework yet addresses deformable environments.Deformable-scene work is hindered by complex or unavailable simulators and difficulty extending current map representations, despite active SLAM’s potential for accurate trajectory estimates.Source: D. Reasoning in Dynamic and Deformable… · pp. 13–14

E. Towards Meaningful and Autonomous Stopping Criteria

• Active SLAM lacks meaningful autonomous stopping criteria: traditional temporal or spatial thresholds do not establish completion in truly unknown environments.TOED-based metrics may identify when exploration stops adding information, but guaranteeing completion, transitioning strategies, and deciding when to stop DRL training remain open questions.Source: E. Towards Meaningful and Autonomous… · p. 14

F. Reproducible Research in Active SLAM

• Active perception still lacks adequate benchmarks, and generalizing evaluations from simulation to the real world and measuring map-estimate quality remain open challenges.Open datasets, simulators, and frameworks offer promise, but commonly used exploration-time and coverage metrics do not directly measure active SLAM’s main objective: improving estimates.Source: F. Reproducible Research in Active SLAM · p. 14

G. Practical Applications

• Active SLAM has few reported field deployments and remains in early development, while computational demands conflict with constraints such as aerial robots’ limited budgets.By 2022, patent searches found about 39,000 for SLAM and 31 for active SLAM; reported applications include underwater inspection and exploration, mines, catacombs, Mars-analogue settings, and assistive mapping.Source: G. Practical Applications · p. 14

• The survey aims to help researchers and practitioners address literature fragmentation and advance active SLAM toward mature algorithms and real applications.It responds to more than three decades of research and renewed attention, including opportunities from learning-based methods.Source: §X · p. 15
• The review presents a unified POMDP formulation, covers goal identification, utility computation, and action selection, and surveys continuous BSP, learning-based, and multi-robot approaches.It also discusses common assumptions, historical evolution, current trends, and state-of-the-art techniques across the modular stages.Source: §X · p. 15
• Open challenges include prediction beyond line-of-sight and active spatial perception, alongside unified formulations, comparable evaluation metrics, reproducibility, and benchmarking.The survey presents these needs as important for the field to mature and achieve real-world impact.Source: §X · p. 15