PaperlayerView exact revision on arXiv

Source-linked AI summary

RAVEN: A Unified Active SLAM System for Rapid Autonomous Visual Exploration with Aerial Robots under Localization Uncertainty

Xinyi Chen, Yichen Zhang, Chenxing Jiang, Shaojie Shen

arXiv:2610.37111v1cs.RO

TL;DR

Aerial active SLAM must balance exploration with localization accuracy, while high agility can aggravate drift and task priorities vary. RAVEN plans tours and perception-aware trajectories using separate exploration and localization utilities. Benchmarks report a trade-off spectrum, with RAVEN and FALCON defining the Pareto front across four scenarios.

  • Problem

    Aerial robots’ agility can exacerbate localization drift, while existing trade-offs may require manual tuning and fail to reflect users’ priorities between exploration efficiency and localization accuracy.

    Source: §I · pp. 1–2

  • Method

    RAVEN generates tours from frontier and active-loop-closure candidates, evaluates separate exploration and localization utilities by user prioritization, and uses perception-aware planning with an online Fisher information field.

    Source: §I · p. 2

  • Results

    RAVEN and FALCON define the Pareto front in exploration time and ATE across all four scenarios, while increasing λ raises exploration time and lowers ATE.

    Source: 2) Result Discussions · p. 13, A. Ablation Study on λ and… · p. 14

  • Takeaways & Limitations

    RAVEN’s prioritization parameter lets it adapt the exploration-efficiency and localization-accuracy trade-off to different task requirements.

    Source: A. Ablation Study on λ and… · p. 14

Abstract

from arXiv · show

Active simultaneous localization and mapping (SLAM) addresses the problem of planning robot motion to build complete and accurate maps of unknown environments through efficient exploration while maintaining high localization accuracy. This raises a dilemma between exploring unknown spaces to expand coverage and exploiting visited regions to perform active loop closures. Despite numerous prior works on ground robots, aerial robots with high agility and constrained sensor suites face more significant localization drift and thereby demand dedicated active SLAM planning. Moreover, a unified strategy that adapts performance based on task-specific priorities remains an open question. In this article, we introduce RAVEN, a unified active SLAM framework for Rapid Autonomous Visual ExploratioN using quadrotors, which systematically accounts for localization uncertainty. The framework starts with identifying frontiers and active loop closure candidates, from which global tour candidates are generated. These candidates are evaluated using a utility vector that separately quantifies exploration efficiency and localization accuracy, and the final tour is selected according to a prioritization parameter ranging from $[0,1]$. Finally, a perception-aware navigation planner generates smooth execution trajectories that minimize visual state estimation error based on the online Fisher information field that provides visual localization information in real-time. Benchmark evaluations demonstrate that RAVEN outperforms state-of-the-art active SLAM methods, highlighting its advantages in producing a spectrum of performance across prioritizations. Fully onboard field experiments with a customized stereo-inertial quadrotor further validate its practical applicability in challenging environments. The source code will be released to benefit the community.

I. INTRODUCTION

RAVEN addresses active SLAM’s exploration–exploitation dilemma for aerial robots by accounting for localization uncertainty across planning. Its related work spans perception-aware planning, classic exploration, and active SLAM.

I. INTRODUCTION

  • Aerial agility can exacerbate localization drift, degrading map quality and increasing loop-closure failure risk; RAVEN addresses uncertainty throughout goal identification, utility evaluation, and action execution.Its global tours balance expected exploration duration and pose-graph information according to user-defined prioritization.Source: §I · p. 2

II. RELATED WORK

  • The related-work review groups prior perceptual planning by localization, mapping, or joint goals, and surveys perception-aware planning, classic exploration, and active SLAM.Source: §II · p. 2

A. Perception-Aware Path Planning

  • Prior perception-aware planners mainly target known environments with known visual features; RAVEN extends this planning scope to exploration without prior environmental knowledge.Earlier approaches use metrics including visibility, feature covisibility, and Fisher information for navigation or trajectory planning.Source: A. Perception-Aware Path Planning · pp. 2–3

B. Classic Autonomous Exploration

1) Classical Methods:

Prior exploration and active SLAM methods use sampling, frontiers, hybrid global guidance, heuristic decisions, information-theoretic utilities, and optimality criteria. Their trade-offs include computation, manual tuning, user preference, and planning horizon.

1) Classical Methods:

  • Classical exploration includes sampling-based viewpoint selection and frontier-based targeting, which trades explicit information gain for computational efficiency.Source: 1) Classical Methods · p. 3

2) Hybrid Methods:

  • Sampling-based exploration selects viewpoints by information gain, while frontier methods efficiently identify informative regions; hybrid methods combine these strengths and often use global guidance.Source: 2) Hybrid Methods · p. 3, 1) Classical Methods · p. 3
  • Hybrid exploration methods combine frontier and sampling strategies, with approaches generating global tours or coverage paths before local refinement or exploration.Despite improved exploration efficiency, methods that omit localization consideration can produce drifted estimates and distorted maps.Source: 2) Hybrid Methods · p. 3

C. Active SLAM

  • Active SLAM unifies exploration and localization goals, evaluating actions through utility functions; the literature review organizes methods by utility type.Source: C. Active SLAM · p. 3

1) Heuristic Decision-Making:

  • Heuristic active-loop-closure strategies switch between exploration and revisiting using uncertainty thresholds, but those thresholds require manual tuning and may not adapt across environments.Source: 1) Heuristic Decision-Making · pp. 3–4

2) Information Theory:

  • Information-theoretic utilities assess information gain from uncertainty in the joint vehicle-pose and map belief, but automatic trade-offs can remain implicit and omit user preference.Entropy-scale differences can require manual weight tuning; one proposed alternative is limited to particle-filter SLAM.Source: 2) Information Theory · p. 4

3) Theory of Optimal Experiment Design (TOED):

  • Optimal experiment design compares actions using covariance-based criteria; prior active SLAM work includes D-optimality and T-optimality measures.Many reviewed methods still select targets greedily, neglecting global exploration-space information.Source: 3) Theory of Optimal Experiment Design… · p. 4

A. Optimality Criteria Based on TOED

B. Graph Theory

The paper uses graph representations to describe connectivity and pose-graph measurements, then approximates pose-graph uncertainty for active SLAM. D-optimality is selected for its monotonicity.

B. Graph Theory

  • An undirected graph’s incidence matrix encodes vertex-edge relationships; its Laplacian and weighted Laplacian provide graph representations used in the subsequent uncertainty discussion.The reduced Laplacian is obtained by removing an arbitrary row and column.Source: B. Graph Theory · p. 5

C. Uncertainty Quantification for a Pose Graph

  • A pose graph represents robot poses as vertices and relative-pose measurements with covariances as edges; full-graph covariance or Fisher information becomes intractable as graph size grows.A graph-topology-based approximation makes D-optimality of the graph Fisher information tractable, and this work uses D-optimality for its monotonicity property.Source: C. Uncertainty Quantification for a… · p. 5

IV. SYSTEM OVERVIEW

Active SLAM plans actions to cover an unknown 3D space while maintaining localization accuracy. RAVEN builds tours from exploration frontiers and loop-closure candidates, then evaluates and executes them through a replanning workflow.

A. Problem Definition

  • The active SLAM objective is to cover an unknown, bounded 3D task space while maintaining localization accuracy and building a complete, accurate map of accessible space.The problem is formulated as a POMDP over map and pose state.Source: A. Problem Definition · p. 5

B. Proposed Framework Overview

  • RAVEN’s workflow identifies vantage points, evaluates global tours for exploration efficiency and localization accuracy, and executes smooth, safe trajectories that minimize visual-odometry drift.The system replans after frontier progress, trajectory completion, or a periodic timer, and stops when no frontier remains.Source: B. Proposed Framework Overview · p. 6

A. Condensed Pose Graph for Visual SLAM

  • The Condensed Pose Graph compresses high-frequency visual state estimates into a sparse graph, reducing computational complexity while preserving information for loop-closure selection and localization-utility evaluation.Its nodes are camera poses; edge covariances summarize feature-reprojection error across intermediate keyframes, and loop closures add edges between nearby nodes.Source: A. Condensed Pose Graph for Visual SLAM · p. 6

B. Vantage Points Selection

  • Exploration viewpoints come from frontiers, while active loop-closure candidates are selected from historical poses whose similar-frame recalls are concentrated.The method clusters CPG nodes by feature co-visibility, then uses NetVLAD similarity to retrieve frames and selects representatives with concentrated recalls.Source: B. Vantage Points Selection · p. 6, Figure 3 · p. 7

C. Global Tour Candidates Generation

  • Rather than greedily selecting the next goal, RAVEN plans global tours across all remaining zones of interest to account for the exploration process through completion.The coverage path visits zone centers; zones and paths are updated from the latest voxel map as exploration progresses.Source: C. Global Tour Candidates Generation · p. 7
  • RAVEN first generates top Ncp coverage-path base tours, then inserts at most one upcoming active loop-closure candidate into eligible tour edges to form a manageable candidate set.The base tours use distance costs derived from info-A* search.Source: C. Global Tour Candidates Generation · p. 7

VI. ACTION EVALUATION WITH PARETO OPTIMALITY

RAVEN evaluates global tours on separate exploration-efficiency and localization-accuracy objectives. It selects among Pareto-optimal tours according to a user-defined prioritization parameter.

VI. ACTION EVALUATION WITH PARETO OPTIMALITY

  • The two objectives are evaluated separately rather than combined into a single utility function, and Pareto order identifies tours that are not dominated on both objectives.Exploration cost is minimized while localization accuracy utility is maximized.Source: §VI · p. 7, C. Global Tour Selection with Pareto… · p. 8

A. Exploration Efficiency Cost

  • Exploration efficiency cost is expected task-completion duration, computed from cumulative traversal time between consecutive tour elements.For a tour with an inserted loop-closure candidate, the expected duration accounts for the detour between zones and that candidate.Source: A. Exploration Efficiency Cost · p. 7

B. Localization Accuracy Utility

  • Localization accuracy utility is expected graph information over virtual CPG configurations representing successful and failed active loop closures.The success probability estimates whether state-estimation deviation remains within acceptable thresholds for revisiting the historical pose; graph information is approximated by D-opt(Lγ(·)).Source: B. Localization Accuracy Utility · pp. 7–8

C. Global Tour Selection with Pareto Optimality

  • The final tour is selected from Pareto-optimal candidates by choosing the point nearest the interpolated position on the Pareto front for the given prioritization parameter λ.This parameter explicitly trades off exploration efficiency against localization accuracy.Source: C. Global Tour Selection with Pareto… · p. 8, Figure 4 · p. 8

VII. PERCEPTION-AWARE ACTION EXECUTION

Perception-aware execution uses an online Fisher Information Field (OFIF) to assess visual localization information and guide trajectories that reduce visual state-estimation drift. Its single-feature information formulation uses right perturbations to avoid dependence on world-frame coordinates while retaining rotation-invariant T-optimality.

VII. PERCEPTION-AWARE ACTION EXECUTION

  • The local planner uses an Online Fisher Information Field to quantify localization information and optimize smooth trajectories that minimize visual state-estimation drift during navigation.The field supports real-time exploration planning with constant-time Fisher-information queries.Source: §VII · p. 8, A. The Online Fisher Information Field · p. 9

A. The Online Fisher Information Field

  • OFIF evaluates how well a 6-DoF camera pose can be estimated from nearby mapped visual features, using a rotation-independent component stored in each voxel and recovered at query time.The total Fisher information sums contributions from visible features; the visibility function indicates whether each feature lies within the camera’s field of view.Source: A. The Online Fisher Information Field · p. 9

1) Fisher Information Matrix for a Single Feature:

  • Right-perturbation FIMs depend on camera-frame feature coordinates, avoiding the world-position-dependent information values of left perturbations while preserving rotation-invariant T-optimality.The comparison concerns features with identical camera-frame coordinates at different world-frame positions.Source: 1) Fisher Information Matrix for a… · p. 9

2) Visibility Model Separation:

The visibility model separates feature visibility into positional and rotational components for OFIF construction. It replaces a binary field-of-view test with a smooth frustum model and uses quaternion-input Gaussian process regression to support that separation.

3) Visibility Model Separation:

  • The original visibility function is binary, but the method approximates it with a smooth, deductive frustum model that accommodates asymmetric fields of view unlike the prior cone model.The frustum model is visualized in Fig. 5, where grey volumes indicate deducted components.Source: 2) Visibility Model Separation · p. 10
  • The method decomposes frustum visibility into positional visibility and rotational visibility, as required for the OFIF factorization.Positional visibility depends on camera and feature positions, while rotational visibility depends on camera rotation.Source: 2) Visibility Model Separation · p. 10
  • Quaternion-input Gaussian process regression approximates visibility across camera rotations using shared rotation samples for each feature and camera position.The regression kernel is a quaternion-adapted squared exponential kernel with learned hyperparameters.Source: 2) Visibility Model Separation · p. 10

3) OFIF Incremental Construction:

OFIF is incrementally constructed from localizable visual features so its free-space voxels retain positional factors for fast information queries. The construction uses a two-step update and approximations that trade FIM accuracy for computational and memory efficiency.

3) OFIF Incremental Construction:

  • OFIF stores positional factors in voxels, enabling constant-time Fisher-information queries by combining the retrieved positional factor with rotational visibility at the queried pose.Its differentiable formulation is also suitable for gradient-based motion-planning optimization.Source: 3) OFIF Incremental Construction · p. 10
  • Only consistently observed and matched features with low position uncertainty are treated as localizable; features outside the exploration bounds or beyond the observation range are excluded.The feature observation range is set larger than the sensing range used for standard mappings.Source: 3) OFIF Incremental Construction · p. 10
  • New free voxels are updated with previous localizable features, then new features update nearby free voxels, with occlusion checks before positional factors are additively updated.This two-step process maintains free-region voxels without duplicated or missing information updates.Source: 3) OFIF Incremental Construction · p. 11, Figure 6 · p. 11
  • Trace-based FIM scalarization, separable visibility, and bounded local updates trade FIM accuracy for memory efficiency and real-time computation.The authors state that the resulting field preserves sufficient visual localization information for the planner’s trajectory generation.Source: 3) OFIF Incremental Construction · p. 11

B. Perception-aware Trajectory Planning

The trajectory planner incorporates OFIF perception information into position and yaw planning. It searches for informative paths and viewing directions, then optimizes smooth, collision-free, dynamically feasible trajectories.

B. Perception-aware Trajectory Planning

  • The position and yaw trajectories are optimized independently using quadrotor differential flatness; the position search accounts for perception because later trajectory optimization only performs local refinements.The Fig. 7 caption contrasts an info-A* path through texture-rich regions with a standard A* shortest path through areas with few localizable features.Source: B. Perception-aware Trajectory Planning · p. 11, Figure 7 · p. 11
  • Info-A* searches for an initial position path using geometric distance and information-potential costs, favoring informative regions while maintaining comparable travel distance to the goal.The heuristic remains Euclidean distance to preserve admissibility; a minimum-time B-spline trajectory is then optimized subject to smoothness, collision avoidance, and dynamic feasibility.Source: B. Perception-aware Trajectory Planning · p. 11
  • Yaw planning first searches sampled yaw candidates along position-trajectory knots for a low-information-cost path, then uses that path to initialize B-spline yaw optimization.The initial yaw path is found with Dijkstra’s algorithm to accelerate optimization convergence and avoid local minima.Source: B. Perception-aware Trajectory Planning · pp. 11–12
  • The B-spline yaw objective combines trajectory jerk as a smoothness cost with accumulated information-potential costs at knot points, balanced by the perception weight λp.The information threshold is based on cumulative information from randomly sampled feature points within sensing range.Source: B. Perception-aware Trajectory Planning · pp. 11–12

VIII. BENCHMARK EXPERIMENTS

The benchmark experiments compare RAVEN with state-of-the-art exploration planners using a simulator extended to support active SLAM evaluation.

VIII. BENCHMARK EXPERIMENTS

  • The benchmark evaluates RAVEN against state-of-the-art exploration planners.The simulation experiments use an Intel Ultra 9 275HX CPU, 96GB RAM, and an NVIDIA RTX 5080 Laptop GPU.Source: §VIII · p. 12

1) Simulator Setup:

  • EPEE+ adds real-time multi-camera image rendering, configurable noisy IMU simulation, and visual SLAM integration to the EPEE simulator.These additions simulate visual-state-estimation drift from visual tracking inaccuracies and IMU noise.Source: 1) Simulator Setup · p. 12
  • EPEE+ provides a photorealistic, lightweight simulator tailored to active SLAM benchmarking and reproducible research.Source: 1) Simulator Setup · p. 12

2) Benchmark Candidates and Settings:

Across four simulated scenarios, RAVEN is evaluated against existing planners using exploration and localization metrics. Its prioritization parameter and perception-aware planner support a trade-off between exploration efficiency and localization accuracy.

2) Benchmark Candidates and Settings:

1) Evaluation Metrics:

  • Evaluation measures exploration time, flight distance, ATE, and successfully matched looped image pairs.ATE is the RMSE of Euclidean distance between the estimated and simulator ground-truth trajectories.Source: 1) Evaluation Metrics · p. 12

2) Result Discussions:

  • Across four scenarios, RAVEN and FALCON form the exploration-time/ATE Pareto front, while RAVEN better balances exploration efficiency and localization accuracy than the other active SLAM methods.RAVEN at λ = 0 has FALCON-comparable exploration duration with lower average ATE; λ = 1 yields consistently low ATE, and λ = 0.5 offers a balanced trade-off.Source: 2) Result Discussions · pp. 12–13, Figure 9 · p. 13, Table III · p. 13

A. Ablation Study on λ and Perception-aware Navigation

  • As λ increases, exploration time rises while ATE falls; the perception-agnostic planner has a dominated Pareto front relative to the perception-aware planner.The study tests six λ values from 0 to 1 in increments of 0.2, with five runs per setting in Warehouse.Source: A. Ablation Study on λ and… · p. 14, Figure 14 · p. 14

B. Ablation Study on Loop Closure Probability Weighting

RAVEN’s probability-weighted loop-closure planning, online computation, and prioritization settings are assessed in simulation and real-world flights. The experiments report adaptable exploration-localization trade-offs, while the conclusion scopes the framework to environments with sufficient visual features.

B. Ablation Study on Loop Closure Probability Weighting

  • Across five independent trials at λ = 1.0, probability weighting achieved average final ATE 0.09 m with 150.6 loop pairs, versus 0.45 m with 10.0 pairs for the first strategy and 0.46 m with 71.8 under uniform thresholding.Source: B. Ablation Study on Loop Closure… · p. 15

C. Exploration in Large-Scale Environment

  • In the large-scale exploration, replanning averaged 0.057 s per iteration and memory remained capped at 2.8 GB.The trial covered 759.6 m and ended with a trajectory ATE of 0.664 m; OFIF and CPG updates ran in separate threads periodically.Source: C. Exploration in Large-Scale… · p. 15, Figure 16 · p. 15

X. REAL-WORLD EXPERIMENTS

  • Fully onboard experiments used a customized quadrotor for autonomous planning, mapping, and localization without human intervention or external infrastructure.The three scenes were the Flight Arena, Campus Hallway, and underground Parking Lot.Source: §X · pp. 15–16
  • In three real-world scenes, increasing λ reduced exploration efficiency while improving localization accuracy; λ = 1 reduced average NND error by 50% in Campus Hallway and 62.5% in Parking Lot relative to λ = 0.The reconstructed maps at λ = 0 showed distortions in the two challenging scenes, while λ = 1 produced higher-fidelity maps.Source: Table IV · p. 15, §X · pp. 16–17, Figure 18 · p. 16

XI. CONCLUSIONS

  • The conclusion limits RAVEN’s intended scope to visually tractable environments under normal indoor lighting, excluding scenes with almost no visual features.LiDAR or radar are identified as potential modalities for future work in those extreme conditions.Source: §XI · p. 17

APPENDIX

The appendix discusses alternative optimality expressions, a closed-form FIM expression, and covariance propagation for CPG edges.

A. Alternative Definitions of D-optimality

  • For a connected graph Laplacian, D-optimality remains well-defined even though the Laplacian determinant is zero.The appendix notes that D-optimality does not necessarily directly relate to det Σ.Source: A. Alternative Definitions of… · p. 17

B. Closed-form Expression of FIM

C. Covariance Computation of CPG Edges

  • For consecutive SLAM frames, 3D feature correspondences and relative transformation define reprojection residuals used to propagate residual covariance to relative-transformation covariance.The residual is written as yi = qi − (Rpi + t) = θ∧Rpi + ρ + ϵi.Source: C. Covariance Computation of CPG Edges · p. 18

Related papers

Loading 2610.37111v1…