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

cs.ROOct 27, 2026

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

Authors: Xinyi Chen, Yichen Zhang, Chenxing Jiang, Shaojie Shen

Organizations: Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Hong Kong, China

Abstract

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.

Figures & tables

Fig. 1: The conceptual picture of the mobile robotics field integration and several community terminologies [ 10 , 11 ] . The highlighted block indicates our focus on active SLAM within this broader context.

Fig. 2: System overview of the RAVEN active SLAM framework for autonomous exploration. After updating vehicle pose estimation and standard volumetric maps with onboard sensors data, the online Fisher information field (OFIF) and condensed pose graph (CPG) are incrementally constructed to facilitate responsive planning. The active SLAM planning proceeds through three phases: goal identification, utility evaluation, and action execution. The framework consistently updates and replans based on the latest sensor measurements until exploration concludes.

Fig. 3: Illustration of ALC candidate selection. CPG node clusters are shown in different colors. The green squares and dashed ellipse highlight the similar frames queried by the pink cluster representative and their spatial distribution, while the red ones indicate those from the purple cluster representative. The pink representative, with concentrated frame recalls, is selected as an ALC candidate, whereas the purple one is excluded due to dispersed frame recalls. The right panel displays the first-person views of the numbered frames, where image 1 is visually distinctive, while image 2 and 3 are similar.

Fig. 4: This figure depicts the global tour candidate selection process. (a) The top $N_{\text{cp}}$ coverage paths connecting all zone of interest centers. (b) Global tour candidates generated by inserting an ALC candidate into the coverage paths. (c) The two possible configurations of virtual CPG, where the upper graph $\mathcal{G}^{+}$ and lower graph $\mathcal{G}^{-}$ represent a loop closure success and failure, respectively. (d) The utility space of the global tour candidates, where the yellow polyline represents the Pareto front with Pareto-optimal solutions in purple circles. The pink inverted triangle marks the selected global tour that is nearest to the interpolated point on the Pareto front according to prioritization parameter $\lambda=0.5$ .

Feature Position		Camera Pos.	Left Pert.	Right Pert.
$\mathbf{p}^{\text{c}}$	$\mathbf{p}^{\text{w}}$	$\mathbf{t}_{\text{wc}}$	$\mathrm{Tr}(\mathcal{I}^{L}(\mathbf{p}))$	$\mathrm{Tr}(\mathcal{I}^{R}(\mathbf{p}))$
(1, 0, 0)	(1, 0, 0)	(0, 0, 0)	4.0	4.0
(1, 0, 0)	(1, 0, 10)	(0, 0, 10)	104.0	4.0
(2, 0, 0)	(2, 0, 0)	(0, 0, 0)	2.5	2.5
(2, 0, 0)	(2, 0, 10)	(0, 0, 10)	27.5	2.5

TABLE I: T-optimality of FIM from Different Perturbations

Fig. 5: The figure shows a visualization of the frustum visibility model, where the grey volumes indicate the deducted components.

Fig. 6: The two-step update of the OFIF incremental construction. (a) In step 1, each new free OFIF voxel is updated with previous localizable features within sensing range. (b) In step 2, each new localizable feature is updated into all nearby free OFIF voxels within sensing range.

Fig. 7: A 2D comparison between info-A* search and standard A* search algorithms for trajectory planning. The central rectangle represents an obstacle and the black dots indicate localizable features. The transparent rainbow-colored histogram depicts the information potential cost at each voxel. The info-A* path (dark blue) favors texture-rich regions with low information panelties, whereas the standard A* search path (red) follows the shortest route, traversing regions with few localizable features.

Parameter	Sec.	Notation	Value
CPG node interval	V-A	$d_{\text{cpg}}$	$1.0$ m
Top base tours	V-C	$N_{\text{cp}}$	$5$
Success loop closure thresh.	VI-B	$\delta_{x},\delta_{y},\delta_{z},\delta_{\psi}$	$3$ m, $0.5$ rad
GPR dimension	VII-A2	$N_{s}$	$360$
Feature observation range	VII-A3	$r_{\text{o}}$	$8.0$ m
Sufficient features	VII-B	$N_{\text{pot}}$	$50$

TABLE II: Parameter Settings

Panel of Figure 10 — Fig. 8: An overview of the four testing scenarios for benchmark experiments.

Scene	Method	Exploration Time (s)				ATE (m)				Flight Distance (m)				Looped Images Count
		Avg	Std	Max	Min	Avg	Std	Max	Min	Avg	Std	Max	Min	Avg	Std	Max	Min
Corridor	SRE [ 16 ]	177.3	18.3	202.0	142.4	0.24	0.15	0.54	0.07	175.2	17.0	194.1	141.7	103.7	66.9	254.0	15.0
	REAL [ 19 ]	153.6	23.0	202.3	126.7	0.20	0.07	0.35	0.10	156.9	20.6	192.4	130.2	118.3	84.2	343.0	52.0
	ExplORB [ 22 ]	159.9	17.9	187.8	126.7	0.17	0.06	0.28	0.10	151.7	24.0	178.0	91.6	79.4	31.3	148.0	28.0
	FALCON [ 21 ]	74.8	6.8	80.9	62.0	0.43	0.09	0.56	0.26	78.1	5.8	89.0	67.6	16.8	22.5	68.0	0.0
	RAVEN - 0.0	88.8	8.0	100.6	76.7	0.29	0.09	0.39	0.12	86.4	7.3	97.3	72.0	15.5	17.3	60.0	0.0

TABLE III: Benchmark Result Statistics

Panel of Figure 12 — Fig. 9: Benchmark result plots of exploration time versus ATE, where average values are marked as triangles and $1$ - $\sigma$ confidence ellipses are depicted.

Panel of Figure 13 — Fig. 10: The final executed trajectories and reconstructed volumetric maps of all the benchmark candidates in the Office scenario. The ground truth trajectories are shown in black curves, and detected loop closures are highlighted with green dots.

Fig. 11: The change in ATE throughout the exploration trails in Fig. 10 , where discontinuities indicate loop closures and pose graph rectification.

Fig. 12: Box plot of planning time per iteration for all benchmarked planners in the Warehouse scenario with medians/averages indicated by lines/triangles.

Panel of Figure 16 — Fig. 13: The final executed trajectories and reconstructed maps of RAVEN-0.0 and RAVEN-1.0 in the Apartment scenario.

Fig. 14: Ablation study results for a series of prioritization parameter $\lambda$ values in the Warehouse scenario. Triangle (Plus) markers and filled (hollow) ellipses indicate means and 1- $\sigma$ confidence intervals using a perception-aware (PAW) (perception-agnostic (PAG)) local planner respectively.

Panel of Figure 18 — Fig. 15: Ablation study results on loop closure probability weighting, showing the reconstructed maps, the estimated and ground truth trajectories in red and black curves, as well as the detected loop closures in green dots.

Panel of Figure 19 — Fig. 16: Top: Scene photo of the Forest scenario. Middle: Exploration results from RAVEN - 0.5. Bottom: Computation breakdown of RAVEN modules throughout the exploration.

		Expl. Efficiency				Localization Accuracy
Scene	$\lambda$	Explore Time(s)		Flight Dist.(m)		† ATE/ NND(m)		Looped Images		‡ Graph Info.
		Avg	Std	Avg	Std	Avg	Std	Avg	Std	Avg	Std
Flight Arena	0.0	44.8	5.19	33.6	3.84	0.18	0.06	4.40	6.47	16.1	5.93
	0.5	48.7	6.09	38.2	5.17	0.14	0.05	17.8	17.3	20.1	7.69
	1.0	51.5	4.97	39.3	6.11	0.10	0.02	36.6	27.3	25.2	8.97
Campus Hallway	0.0	61.2	7.78	57.6	7.72	0.28	0.04	23.7	20.9	18.1	6.56

TABLE IV: Real-World Experiments Statistics

Panel of Figure 21 — Fig. 17: The real-world experiment scene photos.

Panel of Figure 22 — Fig. 18: Visualization of the real-world experiment results. The three rows show results in the Flight Arena, Campus Hallway, and Parking Lot from top to down. The three columns correspond to prioritization parameter $\lambda$ of $0.0,0.5,1.0$ from left to right. In the figures, the odometry trajectory is shown in pink, with purple arrows indicating the vehicle’s yaw directions. The reconstructed occupancy grid is shown in grey, and detected loop closures are marked by green dots. The ground truth odometry trajectory is shown as a black curve in the first row, and the ground truth point cloud is represented in light yellow in the second and third rows. For more details on the complete exploration process, please refer to the supplementary video.

Panel of Figure 23 — Fig. 19: Visualization of the OFIF incremental construction process, illustrating the optimal observation directions in black arrows and localizable features in blue dots during exploration of the Parking Lot scene with $\lambda=0.5$ .

Simultaneous Localization and Mapping

Simultaneous localization and mapping estimates a robot’s motion while building a map of an unknown environment.

See all 198 papers

Dream-SLAM: Dreaming the Unseen for Active SLAM in Dynamic Environments
Feb 25, 2026Xiangqi Meng, Pengxu Hou, Zhenjun Zhao +4Simultaneous Localization and Mapping Visual SLAM
Thrust of Robotics and Autonomous Systems, Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · University of Zaragoza, Zaragoza, Spain. · School of Computation, Information and Technology, Technical University of Munich, Munich, Germany. +1
SLIDER: Sparse History-Guided Aerial Robot Target Search using Sliding Local Maps
Jul 12, 2026Xiaolei Hou, Zheng Pan, Hua Lan +6Aerial Robotics Robot Navigation
Evaluation of Monocular SLAM Systems on High-Altitude Nadir UAV Footage
Aug 19, 2026Gašper Spagnolo, Matej Dobrevski, Danijel SkočajVisual SLAM UAV Navigation
Faculty of Computer and Information Science, University of Ljubljana, Večna pot 113, Ljubljana, Slovenia

Feb 25, 2026J/K move · Enter open · S save

Dream-SLAM: Dreaming the Unseen for Active SLAM in Dynamic Environments

Xiangqi Meng, Pengxu Hou, Zhenjun Zhao, Javier Civera+3

Thrust of Robotics and Autonomous Systems, Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · University of Zaragoza, Zaragoza, Spain. · School of Computation, Information and Technology, Technical University of Munich, Munich, Germany.+1

Simultaneous Localization and Mapping Visual SLAM Robot Navigation

Open paper AskPDF arXiv

In addition to the core tasks of simultaneous localization and mapping (SLAM), active SLAM additionally in- volves generating robot actions that enable effective and efficient exploration of unknown environments. However, existing active SLAM pipelines are limited by three main factors. First, they inherit the restrictions of the underlying SLAM modules that they may be using. Second, their motion planning strategies are typically shortsighted and lack long-term vision. Third, most approaches struggle to handle dynamic scenes. To address these limitations, we propose a novel monocular active SLAM method, Dream-SLAM, which is based on dreaming cross-spatio-temporal images and semantically plausible structures of partially observed dynamic environments. The generated cross-spatio-temporal im- ages are fused with real observations to mitigate noise and data incompleteness, leading to more accurate camera pose estimation and a more coherent 3D scene representation. Furthermore, we integrate dreamed and observed scene structures to enable long- horizon planning, producing farsighted trajectories that promote efficient and thorough exploration. Extensive experiments on both public and self-collected datasets demonstrate that Dream-SLAM outperforms state-of-the-art methods in localization accuracy, mapping quality, and exploration efficiency. Source code will be publicly available upon paper acceptance.

Abstract

Simultaneous Localization and Mapping

Explore similar work

Dream-SLAM: Dreaming the Unseen for Active SLAM in Dynamic Environments

SLIDER: Sparse History-Guided Aerial Robot Target Search using Sliding Local Maps

Evaluation of Monocular SLAM Systems on High-Altitude Nadir UAV Footage