PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Julio A. Placed, Jared Strader, Henry Carrillo, Nikolay Atanasov, Vadim Indelman, Luca Carlone, José A. Castellanos

arXiv:2207.00254v2cs.RO

TL;DR

Active SLAM asks how robots can control their motion while localizing and mapping unknown environments accurately and completely. This survey unifies the problem, reviews modular and alternative solution approaches, and identifies open challenges for reproducible research and real-world impact.

  • Problem

    Active SLAM has developed across communities with disparate approaches and formulations, creating a need for a cohesive review and framework.

  • Method

    The paper surveys active SLAM through a unified POMDP formulation, a modular scheme for goal identification, utility computation, and action selection, alternative planning and learning approaches, and multi-robot work.

  • Results

    The survey synthesizes historical developments, current trends, state-of-the-art techniques, and open challenges in active SLAM.

  • Takeaways & Limitations

    Active SLAM research needs unified formulations, evaluation metrics, reproducible methods, and benchmarks to mature and achieve real-world impact.

Abstract

from arXiv · show

Active Simultaneous Localization and Mapping (SLAM) is the problem of planning and controlling the motion of a robot to build the most accurate and complete model of the surrounding environment. Since the first foundational work in active perception appeared, more than three decades ago, this field has received increasing attention across different scientific communities. This has brought about many different approaches and formulations, and makes a review of the current trends necessary and extremely valuable for both new and experienced researchers. In this work, we survey the state-of-the-art in active SLAM and take an in-depth look at the open challenges that still require attention to meet the needs of modern applications. After providing a historical perspective, we present a unified problem formulation and review the well-established modular solution scheme, which decouples the problem into three stages that identify, select, and execute potential navigation actions. We then analyze alternative approaches, including belief-space planning and deep reinforcement learning techniques, and review related work on multi-robot coordination. The manuscript concludes with a discussion of new research directions, addressing reproducible research, active spatial perception, and practical applications, among other topics.

I. INTRODUCTION

Active SLAM jointly controls robot motion, localization, and mapping to build an accurate and complete model while balancing exploration and exploitation. This survey responds to the field’s divergence by reviewing current approaches and open challenges, including reproducibility and benchmarks.

  • Active SLAM jointly resolves localization, mapping, and motion control to create an accurate and complete model of an unknown environment.
  • The decision-making problem balances exploring new areas with exploiting observed regions to improve the resulting map.
  • Research has expanded across communities and approaches, broadening the field while creating a disconnect between research lines.
  • The survey reviews emerging opportunities in spatial perception and artificial intelligence, including prediction beyond line-of-sight and novel environment representations.
  • The paper identifies open challenges for real applications and emphasizes benchmarks and reproducible research.

A. Historical Perspective

Active SLAM emerged by combining mapping, localization, and navigation into autonomous operation in initially unknown environments. Its history spans active mapping, active localization, and active perception, followed by multiple names for the unified problem.

  • Autonomous robots require environmental models, localization, mapping, and safe navigation to perform tasks independently.
  • SLAM incrementally builds an environmental map while simultaneously locating the robot within it.
  • Active perception adds decision-making to data acquisition, enabling navigation actions aimed at improving localization, mapping, or both.
  • Active mapping searches for movements that create the best possible environmental representation, initially assuming perfectly known sensor localization.
  • Active localization determines robot motion to minimize future expected pose uncertainty when the map is known.
  • Active SLAM unifies active localization and mapping by controlling a robot to reduce uncertainty in its pose and map.
  • The unified problem appeared under varied names, including active exploration, autonomous SLAM, belief-space planning, and robotic exploration.

B. About Previous Surveys

Earlier surveys established historical context but left important gaps in unified formulation and coverage. This survey distinguishes itself by addressing utility computation alongside broader active SLAM topics.

  • Only two earlier works had specifically surveyed active SLAM research.
  • Cadena et al. covered history and main aspects but lacked detailed work discussion and a unified mathematical formulation because of brevity.
  • Their survey identified stopping, future-action prediction, and mathematical optimality guarantees as open challenges.
  • Lluvia et al. provided historical and cross-community coverage but did not present a unified mathematical formulation or cover utility computation.
  • Lluvia et al. examined vantage-point and trajectory optimization, which had recently attracted research interest.

Candidate Goal

The survey presents active SLAM through environment representations, probabilistic belief modeling, and a traditional three-stage modular workflow. This workflow reduces the action search space, evaluates candidate utilities, and selects actions, although decoupling can produce suboptimal behavior.

  • Candidate Goal: Belief-space modeling represents probability density functions over system states and uses history comprising observations and past control actions.
  • Candidate Goal: Future-action evaluation requires predicting posterior beliefs and studying expected observations because future measurements are unknown.
  • Candidate Goal: The modular scheme identifies candidate actions, computes their expected costs and gains, then selects and executes optimal actions iteratively.
  • Candidate Goal: Decoupling the stages can produce suboptimal results and undesired behaviors, whereas simultaneous optimization is advantageous for continuous actions and learned POMDP policies.
  • Candidate Goal: The survey compares topological, metric, metric-semantic, and hybrid map representations for active SLAM.

2) Metric maps

Metric maps encode geometric information about the environment and are widely used in active SLAM. They range from sparse landmark representations to dense maps and specialized signed distance fields.

  • 2) Metric maps: Signed distance field maps such as Voxblox encode more sophisticated geometric metrics but are seldom used in active SLAM.

3) Metric-semantic maps

The survey discusses hybrid representations that combine metric, semantic, and topological information to support higher-level reasoning. It also reviews goal-detection strategies that balance exploration with localization and mapping accuracy.

  • 3) Metric-semantic maps: Hybrid maps combine previous representations to enhance decision-making, including metric-topological navigation and SLAM models.
  • 3) Metric-semantic maps: Three-dimensional scene graphs combine metric, semantic, and topological representations into interconnected high-level entities for high-level reasoning.
  • 3) Metric-semantic maps: Enumerating every reachable destination is intractable, so practical methods evaluate finite candidate subsets without guaranteeing global optimality.
  • 3) Metric-semantic maps: Frontier detection evaluates observed free-space regions, with three-dimensional methods less frequent because 3D maps are costly to store and analyze.
  • 3) Metric-semantic maps: Potential loop-closure regions can be more informative than frontiers under high uncertainty, motivating combined candidates or switching between exploration and revisiting.
  • 3) Metric-semantic maps: Local goal identification can be optimal only locally, and short decision horizons may induce wrong behaviors in deep reinforcement learning approaches.
  • 3) Metric-semantic maps: Utility computation evaluates candidate destinations using objectives such as travel cost, mapping uncertainty, and localization uncertainty.
  • 3) Metric-semantic maps: Naive geometric metrics include Euclidean distance, travel time, and expected visited area, with expected area approximating occupancy-grid map entropy.

B. Information Theory (IT)

Information-theoretic active SLAM utilities quantify uncertainty in the joint belief state, commonly through entropy or expected entropy reduction, while practical computation relies on approximations and representation-specific assumptions.

  • Information-theoretic metrics: Entropy measures uncertainty in the joint belief state and underlies information-theoretic utility metrics in active SLAM.The surveyed metrics include entropy, mutual information, and Kullback–Leibler divergence.
  • Approximation and limitations: Joint SLAM entropy is generally intractable, so methods often approximate robot and map utilities independently before combining them heuristically.Additive approximations can require weighting parameters because the robot and map entropy terms have mismatched magnitudes.
  • Representation-dependent entropy: Map entropy depends on the chosen representation; occupancy-grid formulations commonly assume independent cells and may evaluate only nearby cells for efficiency.This measure has been used in both 2D and 3D occupancy-grid maps, with local evaluations proposed for particle-filter SLAM.
  • Information-theoretic metrics: Mutual information measures information gain as the difference between current entropy and expected post-action entropy.The expectation is taken over expected future measurements for each candidate action.
  • Information-theoretic metrics: Kullback–Leibler divergence captures changes in both probability-density shape and mean location between prior and posterior distributions.The distributions may be prior and posterior, or estimated and true posteriors when the latter can be approximated.
  • Approximation and limitations: For occupancy-grid maps, entropy, mutual information, and Kullback–Leibler divergence rely on cell counting, making them discrete and poorly suited to optimization.Differentiable cost-utility functions address this issue for 2D occupancy grids and voxel maps but assume perfect robot localization.

C. Theory of Optimal Experimental Design (TOED)

Optimal experimental design evaluates active SLAM actions through covariance-based uncertainty in task-relevant variables, offering criteria that complement information-theoretic measures but face computational and planning limitations.

  • TOED foundations: TOED-based utilities quantify uncertainty directly in task space through covariance of Gaussian variables of interest.An action set is preferred when the joint posterior covariance is smaller.
  • Optimality criteria: Kiefer’s criteria derive from eigenvalue-based covariance mappings: T-opt captures average variance, D-opt covariance-ellipsoid volume, A-opt harmonic-mean variance, and E-opt ellipsoid radii.These correspond to p = 1, p = 0, p = −1, and p →±∞, respectively.
  • Optimality criteria: E-optimality captures the radii of the covariance hyperellipsoid and corresponds to the limiting case p →±∞.The criterion is defined through the eigenvalues of the covariance matrix.
  • Use in active SLAM: Optimality criteria entered active SLAM through covariance-ellipse area, with later methods focusing mainly on T-optimality and D-optimality while information-theoretic methods remain more popular.Both map and robot uncertainties must be represented by a covariance matrix.
  • Computational considerations: Covariance-based utility evaluation is expensive for large dense matrices, motivating Fisher-information and pose-graph connectivity analyses as computational alternatives.Graph connectivity has been shown equivalent to computing optimality criteria in active graph-SLAM, including links between spanning trees and D- and E-optimality.
  • Planning with TOED utilities: The modular workflow selects an informative destination before path planning, but decoupling destination choice from path computation rarely yields an optimal solution to the original problem.Trajectory-optimization and belief-space-planning methods address this suboptimality by optimizing future motion more directly, although exact continuous-state POMDP solutions are generally intractable.

B. On the Reward Function Design and the Action Set

Deep reinforcement learning for active SLAM uses extrinsic or intrinsic rewards to select exploratory actions, increasingly incorporating uncertainty metrics while retaining unresolved challenges in joint reward design and generalization.

  • Reward design: Early DRL exploration methods learned actions from raw observations using Q-networks and later parallel architectures for more complex unseen scenarios.These approaches initially used purely extrinsic rewards supplied by the environment.
  • Reward design: Intrinsic motivation and curiosity rewards encourage agents to visit unknown configurations, while coverage rewards support exploration in complex 3D simulations.Open-source approaches also report benefits from pre-training and combining multiple input sources.
  • Uncertainty-aware rewards: DRL rewards can incorporate D-optimality, T-optimality, or map mutual information to provide foundations from established uncertainty theories.Current agents using these metrics target either localization or mapping uncertainty, while accounting for both remains open.
  • Learning and planning: Combining learning with planning can make policies easier to learn, improve generalization, and support transfer across platforms.Several methods use DRL to select among frontiers or nearby sampled locations while leaving motion planning outside learning.
  • Partial observability and generalization: Partial observability makes single observations insufficient to identify the robot’s true state, leaving learned policies potentially suboptimal.Using previous observations and recurrent processing are described as ways to address this issue.
  • Partial observability and generalization: Generalization to unseen scenarios is a key limiting factor, with mitigation strategies including randomized starts, observation noise, sparse inputs, egocentric observations, and graph representations.Transfer of trained agents to real scenarios remains an open research problem with few efforts reported.

D. Training Environments

Active SLAM training and deployment span simplified and realistic simulators, multi-robot coordination, and broader informative path-planning objectives, while substantial research remains necessary for complex real-world environments.

  • Simulation environments: DRL training requires online data from simulation because offline supervised training is unavailable and direct real-world training appears infeasible.Simplified scenarios may restrict inputs to ground-truth or perfect range observations, limiting realism.
  • Simulation environments: Stage provides simple 2D bitmap environments, whereas Gazebo supports 3D simulation, realistic rendering, visual sensors, and ROS integration.CoppeliaSim/V-REP additionally supports online mesh manipulation.
  • Simulation environments: AI-oriented simulators prioritize training speed, with DeepMind Lab offering discrete low-texture scenarios and Habitat-Sim supporting physics and varied robot and visual sensor models.These platforms differ from robotics-oriented simulators adapted for DRL.
  • Multi-robot settings: Multi-agent active SLAM jointly optimizes sensing trajectories over combined state and action spaces, but computational complexity becomes infeasible as robot count increases.Collaborative exploration instead coordinates frontier assignments using frontier utility and travel cost.
  • Broader objectives: Informative path planning generalizes active SLAM to objectives beyond localization and mapping, including target tracking and environmental monitoring.Multi-agent methods also evaluate entropy along trajectories and optimize belief-dependent information objectives.
  • Multi-robot settings: Collaborative active estimation selects multi-robot trajectories that reduce localization and map uncertainty using relative-entropy optimization or belief-space planning.These approaches extend active SLAM reasoning to coordinated uncertainty reduction.
  • Open challenges: Supporting fully autonomous robots in complex environments still requires substantial research and cooperation across the field’s involved research areas.The survey identifies both long-standing and comparatively underexplored challenges.

A. Prediction Beyond Line-of-sight

Active SLAM increasingly uses learned scene and map prediction to anticipate information beyond a sensor’s field of view. Integrating these predictions with belief-space planning remains an important challenge, especially for forecasting uncertainty and loop closures.

  • A. Prediction Beyond Line-of-sight: Traditional active SLAM predicts sensed space and maps with sensor models and ray-casting, while recent methods use learned representations for beyond-line-of-sight prediction.Approaches include augmented occupancy-grid prediction, input-map extension, occupancy-probability prediction, and 3D signed-distance-field scene prediction.
  • A. Prediction Beyond Line-of-sight: Auto-encoders and variational auto-encoders predict unseen occupancy or map structure from sensor measurements or existing occupancy-grid maps.Other work predicts 3D scenes using signed-distance-field maps.
  • A. Prediction Beyond Line-of-sight: Integrating fast learned map prediction into active SLAM remains incomplete despite promising results for online prediction beyond the sensor’s field of view.The survey identifies this integration as an open issue rather than an established solution.
  • A. Prediction Beyond Line-of-sight: Predicting robot-state uncertainty is difficult, and few studies combine data-driven scene prediction with belief-space planning.One reported effort also forecasts posterior uncertainty over the robot trajectory from predicted observations.
  • A. Prediction Beyond Line-of-sight: Active SLAM must also account for how robot motion affects future loop-closure observations and beliefs.The passage introduces loop closures as another factor alongside movement and uncertainty prediction.

B. From Active SLAM to Active Spatial Perception

The survey broadens active SLAM from geometric mapping toward semantic, hierarchical, ambiguous, dynamic, and deformable environments. It also identifies unresolved challenges in completion criteria, benchmarking, deployment, and evaluation.

  • B. From Active SLAM to Active Spatial Perception: Active spatial perception extends active SLAM toward metric-semantic and hierarchical representations, including objects, rooms, people, and buildings.The survey notes that most existing approaches still reason over geometric representations such as occupancy-grid maps.
  • B. From Active SLAM to Active Spatial Perception: Semantic knowledge may improve unseen-space prediction and SLAM, while hierarchical representations may support more computationally efficient planning.Examples include predicting rooms or objects and using semantics for new loop-closure detection methods.
  • B. From Active SLAM to Active Spatial Perception: Open questions include uncertainty quantification, computation, and spatial prediction for metric-semantic and hierarchical scene representations.These questions limit the current use of active spatial perception in active SLAM.
  • C. Robust Online Belief Space Planning and Active SLAM: Ambiguous data association can cause catastrophic failures, motivating active SLAM methods that plan to disambiguate future association hypotheses.Recent belief-space planning work explicitly reasons about future data-association hypotheses.
  • D. Reasoning in Dynamic and Deformable Scenes: Dynamic and deformable environments remain underdeveloped because active SLAM typically assumes static scenes and has largely restricted dynamic-scene reasoning to replanning.No deformable active SLAM framework had been developed in the surveyed literature, partly because suitable simulators and map representations are difficult to construct.
  • E. When Is Active SLAM Complete?: Active SLAM completion is difficult to determine because online utility estimation and optimization consume substantial computational resources.A completion criterion is needed to decide when other tasks can be prioritized.
  • F. Reproducible Research in Active SLAM: Active SLAM needs reproducible benchmarks and metrics beyond exploration time and coverage because interactive experiments hinder direct comparisons.Simulation-to-real generalization and measuring estimate quality remain open evaluation questions.
  • G. Practical Applications of Active SLAM: Despite practical relevance, few active SLAM systems have been deployed, and patent counts indicate technology readiness remains in early development.The survey reports about 39,000 patents using “SLAM” versus 31 using “active SLAM” by 2022.

X. CONCLUSIONS

The survey responds to fragmentation in active SLAM by organizing its history, formulations, methods, and open challenges into a cohesive framework. It emphasizes unified evaluation and reproducibility as prerequisites for field maturation and real-world impact.

  • X. CONCLUSIONS: Active SLAM controls a robot to estimate an accurate and complete environmental model, but disparate literature has hindered cohesive frameworks and real-world transition.The survey positions itself as a guide for researchers and practitioners addressing this fragmentation.
  • X. CONCLUSIONS: The paper formulates active SLAM under POMDPs and reviews a modular scheme of goal identification, utility computation, and action selection.It also surveys continuous belief-space planning, learning-based approaches, and multi-robot active SLAM.
  • X. CONCLUSIONS: The survey identifies prediction beyond line-of-sight and active spatial perception among the field’s major open challenges.It also highlights the need for unified formulations and evaluation metrics supporting direct comparison.
  • X. CONCLUSIONS: Reproducibility and benchmarking must be addressed for active SLAM to mature and achieve real-world impact.The conclusion explicitly connects these needs with field maturation and practical impact.
Loading 2207.00254v2…