BatSLAM 2.0: Sequence-Verified Sonar Place Recognition in a Robust Pose Graph

Jan Steckel

cs.ROSep 30, 2026

BatSLAM 2.0: Sequence-Verified Sonar Place Recognition in a Robust Pose Graph

Authors: Jan Steckel

Organizations: Cosys-Lab, Faculty of Applied Engineering, University of Antwerp, 2020 Antwerp, Belgium · Flanders Make Strategic Research Centre, 3920 Lommel, Belgium

Abstract

Echolocating bats can navigate dark and cluttered spaces using echolocation. Over a decade ago, BatSLAM showed that a robot with a biomimetic binaural sonar can build a topological map of the environment, by recognizing places from the received acoustic signals. Sonar place recognition, however, is ambiguous by nature: corridors produce nearly identical echo trains, and wrong loop closure can collapse the topological map. In this paper, we introduce BatSLAM 2.0, a novel sonar-only SLAM system built from three elements: an updated acoustic front-end, a sequence verifier that tracks and verifies loop closure candidates and a pose graph implemented on a high performance factor graph framework. The system was thoroughly evaluated both in simulated as well as real world recordings. In both cases, the BatSLAM2.0 algorithm shows the capability of robust topological map creation, countering map collapse, and robust scaling of map size.

Figures & tables

Panel of Figure 1 — Fig. 1: Overview of the processing flow of BatSLAM 2.0. For every pulse, the binaural echo is converted into a local view $V$ with an energy image $E$ and a spectral-shape image $S$ (the direction cue), which is compared with all stored templates. The five best matches are candidates for the sequence verifier, which keeps several recognition hypotheses alive and commits to one only when it is long, strong, unambiguous and plausible given the current pose uncertainty. A committed hypothesis injects weak loop closure links into the pose graph, which is solved incrementally with iSAM2. The links of every hypothesis form a group that the link management can withdraw from the graph or re-admit later. New templates are anchored to pose-graph nodes, so that their poses follow every optimization, and no templates are laid while the robot is locked onto known territory.

Fig. 2: Processing steps of the acoustic front-end, illustrated on pulse 2600 of the development drive on the indoor floor. Panel a) shows the received binaural echo train, normalized per ear, as a function of range. Panel b) shows the output of the matched filter for the left ear in decibels, which reveals weak echoes up to $10\text{\,}\mathrm{m}$ that are invisible in the raw signal. Panel c) shows the fine cochleogram (64 bands between $30\text{\,}\mathrm{kHz}$ and $90\text{\,}\mathrm{kHz}$ , $3\text{\,}\mathrm{cm}$ range bins) as linear amplitude, for both ears (bands low to high within each ear), and panel d) shows the same cochleogram after the time-varying gain. Panel e) shows the energy image $E$ of the descriptor (32 pooled bands, $6\text{\,}\mathrm{cm}$ range bins, cube-root compression), and panel f) the spectral-shape image $S$ (red: bands above the mean level of the range bin, blue: below), which carries the direction-dependent filtering of the ears and the emitter. The speckle in panel f) beyond $5\text{\,}\mathrm{m}$ is noise that passes the level mask.

Fig. 3: Snapshots of five places and their local views. The top row shows the scene within

5\text{\,}\mathrm{m}

of the robot (circle, with an arrow for its heading), with the frontal half-space of the sonar shaded in blue and the complete route in light grey. The middle row shows the energy image

E

and the bottom row the spectral-shape image

S

of the local view (red: above the mean level of the range bin, blue: below), each with the left ear above the right ear. Panels a) and b) show the same place in a corridor of the indoor floor, visited

641\text{\,}\mathrm{m}

of travel apart; panel c) shows the view of another place,

7.7\text{\,}\mathrm{m}

away, that is most similar to b) among all earlier pulses of the drive; panel d) shows the pillar hall of the indoor floor, and panel e) a street in the city. The similarity

s

is 0.89 between a) and b), and 0.69 between b) and c).

Fig. 4: Directivity of the simulated sonar (Lambert azimuthal equal-area projection of the frontal hemisphere, seen from behind the head; grid lines every $30\text{\,}\mathrm{\SIUnitSymbolDegree}$ ). Rows: emitter, left ear, right ear and the ERTFs of both ears, at five frequencies between $30\text{\,}\mathrm{kHz}$ and $90\text{\,}\mathrm{kHz}$ , each normalized to its maximum.

Fig. 5: The simulated worlds and exemplary results of BatSLAM 2.0. Top row: the ground-truth drive (colored from start to end) in a) the indoor floor (

38\text{\,}\mathrm{m}

18\text{\,}\mathrm{m}

, development drive on route 3,

964\text{\,}\mathrm{m}

), b) the city (

46\text{\,}\mathrm{m}

26\text{\,}\mathrm{m}

) and c) the small development hall (

16\text{\,}\mathrm{m}

10\text{\,}\mathrm{m}

). Black crosses: the two copies of the aliasing pattern. Bottom row, d)–f): the final map estimated by BatSLAM 2.0 for the same drives with calibrated odometry (blue), with its trajectory error (ATE); wrong links in the final graph are drawn in red.

Fig. 6: Single-view place recognition for four sensors on the city drive: a) top-1 accuracy, b) AUC. Grey: descriptor of the original BatSLAM; orange: the same with TVG; blue: the final descriptor.

Fig. 7: Capture region of local view templates on the indoor floor, averaged over 16 template poses: fraction of recognized offset views for combined a) along-track and heading offsets, and b) lateral and heading offsets.

Fig. 8: Loop closure matrix of the development drive (route 3, calibrated odometry). Every dot is a link in the final graph between a query pulse and the anchor of the matched template. Grey: true same-direction revisits; blue: correct links; red: wrong links.

Fig. 9: Robustness against systematic odometry errors on the development drive. a) Trajectory error of odometry (orange), BatSLAM 2.0 (blue) and BatSLAM 2.0 after alignment (green) versus the bias factor

b

; b) wrong hypotheses in the final graph and percentage of collapsed pairs.

Fig. 10: The real-world recording. a) Occupancy grid of the hall from the lidar scans, with the ground-truth trajectory (colored by time); the left loop was driven twice in the same direction, the right loop once in each direction. b) Camera images at places A, B and C. c) Binaural cochleograms of the closest sonar pulse, formed with the virtual ears of equation 16 .

Fig. 11: Real-world results. a) Map of one run (refitted evidence curve) with odometry, ground truth, BatSLAM 2.0 estimate and links. b) Similarity of correct and wrong candidates, with the simulated (dashed) and refitted (solid) evidence curves.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Parameter	Value
Front-end	Filterbank	64 Gaussian bands, $30\text{\,}\mathrm{kHz}90\text{\,}\mathrm{kHz}$ , log-spaced, width of two band spacings
	Band pooling	Pairs of bands (32 bands)
	Range bins	$6\text{\,}\mathrm{cm}$ , from $0.25\text{\,}\mathrm{m}$ to $10\text{\,}\mathrm{m}$ (162 bins)
	Time-varying gain	Amplitude $\times r$
	Energy image	Cube root of the amplitude, relative to the peak of the view
	Spectral-shape image	$\pm$ $15\text{\,}\mathrm{dB}$ scaled to $\pm 1$ , only above $-40\text{\,}\mathrm{dB}$

Appendix

TABLE I: Parameters of BatSLAM 2.0. The same values were used for every experiment in this paper.

Cues
	Indoor floor		City
Variant	Top-1	AUC	Top-1	AUC
Energy	0.875	0.915	0.878	0.918
Energy + shape (ours)	0.913	0.945	0.888	0.938
Energy + shape + ILD	0.892	0.968	0.878	0.960
Compression of the energy image (energy + shape)

Appendix

TABLE II: Single-view place recognition for the design choices of the descriptor, with the design sensor. All variants use 32 channels,

6\text{\,}\mathrm{cm}

range bins and the TVG. Our descriptor uses a cube-root compression. Best value per column in bold. ILD: interaural level difference.

Condition (runs)	Odometry	Ours	Aligned	Wrong	Coverage
	ATE [m]
Calibrated (6)	4.2–11.1	0.47–1.91	0.18–0.46	0/230	0.79–0.84
Residual bias (6)	11.0–19.0	0.62–1.65	0.23–0.53	0/227	0.79–0.84
Degraded sensors (3)	7.5	0.52–1.79	0.25–0.81	1/106	0.67–0.74
Uncalibrated (4)	19.1–22.3	1.38–12.68	0.45–2.96	0/149	0.62–0.81

Appendix

TABLE III: Results of BatSLAM 2.0 per condition (ranges over the runs). ATE: trajectory error of odometry and of BatSLAM 2.0, without and after rigid alignment; wrong: wrong committed hypotheses in the final graph, out of all committed hypotheses; coverage: fraction of true revisits with a correct link.

Variant	ATE [m]	Wrong	Coverage
Full system (ours)	0.87	0	0.81
Naive	18.90	364	0.72
No sequence verification	4.85	10	0.80
Original front-end (8 bands)	1.54	0	0.73
No time-varying gain	2.53	3	0.72
Energy image only	1.38	8	0.74

Appendix

TABLE IV: Ablations over four standard cases (routes 3 and 5 with calibrated odometry, route 7 with a residual bias, and the city). ATE: mean over the cases; wrong: wrong committed hypotheses in the final graphs, summed over the cases. Naive: no sequence verification and no link management.

Evidence curve	ATE	Aligned	Links	Precision	Wrong	Revisits	Collapsed
	( $\mathrm{m}$ )	( $\mathrm{m}$ )			Hyp.	Linked	Pairs
Odometry only	3.62 (0.46)
Simulation evidence curve	0.77 (0.07)	0.60 (0.05)	108	1.000 (0.000)	0	0.50 (0.01)	0.000
Refitted, even stretches	0.67 (0.13)	0.52 (0.11)	141	0.980 (0.018)	0	0.64 (0.10)	0.000
Refitted, odd stretches	0.66 (0.12)	0.52 (0.11)	142	0.975 (0.021)	0	0.64 (0.10)	0.000

Appendix

TABLE V: Real-world results, mean (standard deviation) over ten odometry seeds (distance

\times 0.95

, rotation

\times 1.10

Fig. 12: Trajectories of all runs of the main experiment (table VI ), sorted by the odometry condition. Grey: ground truth; orange: odometry; blue: BatSLAM 2.0. Run names: r3 , r5 , r7 : routes 3, 5 and 7 on the indoor floor; city : the city world; b0 , b025 , b1 : bias factor $b=0$ , $0.25$ and $1$ ; s1 – s3 : odometry noise realization; noise5e-6 , noise1.5e-5 , oldsensor : degraded sensors.

Fig. 13: Trajectories of all runs of the robustness experiment (table VII ). bias $b$ _s $k$ : bias factor $b$ with odometry realization $k$ ; odomnoise_x3 : three times the random odometry noise; the last three runs use degraded sensors with a residual bias. Colors as in figure 12 .

Case	Odometry	ATE odo	ATE	Aligned	Precision	Wrong (final)	Wrong (all)	Coverage	Collapsed
		[m] $\downarrow$	[m] $\downarrow$	[m] $\downarrow$	$\uparrow$	$\downarrow$	$\downarrow$	$\uparrow$	[%] $\downarrow$
Indoor, route 3, seed 1	Calibrated	7.5	0.78	0.25	0.999	0/39	0	0.84	0.0
Indoor, route 3, seed 2	Calibrated	4.2	0.89	0.24	0.999	0/40	0	0.84	0.0
Indoor, route 3, seed 3	Calibrated	7.9	1.02	0.19	1.000	0/36	1	0.84	0.0
Indoor, route 5	Calibrated	10.1	0.47	0.23	0.998	0/45	0	0.82	0.0
Indoor, route 7	Calibrated	7.6	1.91	0.18	1.000	0/46	0	0.80	0.0

Appendix

TABLE VI: Results of BatSLAM 2.0 on all runs of the main experiment (drives of about

1\text{\,}\mathrm{km}

). ATE odo: dead reckoning; ATE and aligned: BatSLAM 2.0 without and after rigid alignment; precision: fraction of correct links; wrong (final / all): wrong committed hypotheses in the final graph and over all commits; coverage: fraction of true revisits with a correct link; collapsed: fraction of collapsed pairs.

Case	Bias	ATE odo	ATE	Aligned	Precision	Wrong (final)	Wrong (all)	Coverage	Collapsed
		[m] $\downarrow$	[m] $\downarrow$	[m] $\downarrow$	$\uparrow$	$\downarrow$	$\downarrow$	$\uparrow$	[%] $\downarrow$
Indoor, route 3, seed 4	$b=0$	8.0	1.22	0.30	1.000	0/38	1	0.82	0.0
Indoor, route 3, seed 5	$b=0$	7.6	0.34	0.20	0.999	0/40	1	0.83	0.0
Indoor, route 3, seed 4	$b=0.25$	15.9	1.98	0.33	0.999	0/38	1	0.83	0.0
Indoor, route 3, seed 5	$b=0.25$	20.7	0.45	0.24	0.999	0/37	2	0.83	0.0
Indoor, route 3, sensor $-20$ dB	$b=0.25$	11.0	1.98	1.43	0.988	1/33	3	0.65	0.2

Appendix

TABLE VII: Robustness runs: odometry bias factor

b

, three times the random odometry noise, and degraded sensors with a residual bias. Columns as in table VI .

Variant	$-10$ dB sensor	$-20$ dB sensor	Original sensor	Route 3, seed 3	Wrong (final)	Wrong (all)
No risk-scaled commit	0.26 (0)	0.24 (0)	1.06 (1)	0.19 (0)	1	10
Risk-scaled commit, length only (24 pairs)	0.26 (0)	0.24 (0)	0.25 (0)	0.19 (0)	0	5
Plausibility gate on the absolute marginal	0.26 (0)	0.23 (0)	0.69 (1)	0.19 (0)	1	12
No plausibility gate	0.26 (0)	1.37 (2)	0.60 (1)	0.19 (0)	3	19
No length rule	0.26 (0)	0.81 (2)	1.07 (3)	0.20 (2)	7	7
No link management	0.26 (0)	0.81 (2)	1.07 (3)	0.20 (2)	7	7

Appendix

TABLE VIII: Ablation of the safeguards on the four hardest cases (three degraded sensors and route 3 with odometry seed 3). Every cell lists the aligned trajectory error (m) and, in parentheses, the number of wrong hypotheses in the final graph; the last two columns sum the wrong hypotheses in the final graphs and over all commits.

Variant	Route 3	Route 5	Route 7, residual bias	City
Naive (single views, no gates, no link mgmt.)	8.00 (63)	26.70 (68)	8.22 (72)	32.68 (161)
No sequence verification (single views)	5.08 (1)	1.95 (1)	8.17 (6)	4.21 (2)
Original front-end (8 wide bands)	1.97 (0)	0.44 (0)	3.15 (0)	0.59 (0)
No time-varying gain	0.85 (1)	6.82 (2)	1.84 (0)	0.62 (0)
Energy image only (no shape)	0.63 (0)	1.70 (3)	1.26 (4)	1.92 (1)
No range-shift search	0.45 (0)	1.57 (0)	3.28 (0)	0.56 (0)

Appendix

TABLE IX: Ablations per case: trajectory error (ATE, in m) and the number of wrong committed hypotheses in the final graph (in parentheses), for route 3 and route 5 with calibrated odometry, route 7 with a residual bias, and the city with calibrated odometry.

Explore similar work

SOL-SLAM: Inverse Compositional Gauss-Newton Direct Registration for Fast Sonar-Only Local SLAM
Sep 16, 2026Kalvik Jakkala, Jason O'KaneForward-Looking Sonar Simultaneous Localization And Mapping
Texas A&M University
SLAM in Low-Light Environments: Project Report
Jul 20, 2026Oleh Basystyi, Anna Stasyshyn, Oleksandr Kosovan +1Orb-Slam2 Simultaneous Localization And Mapping
Applied Sciences Faculty, Ukrainian Catholic University
SLAMSqueezeBench: Comparing SLAM Systems under Resource Constraints
Sep 17, 2026Mohamed Hefny, Karthik Dantu, Steven Y. KoSimultaneous Localization And Mapping
Simon Fraser University. · University at Buffalo.

Sep 16, 2026, cs.ROJ/K move · Enter open · S save

SOL-SLAM: Inverse Compositional Gauss-Newton Direct Registration for Fast Sonar-Only Local SLAM

Kalvik Jakkala, Jason O'Kane

Texas A&M University

Forward-Looking Sonar Simultaneous Localization And Mapping Odometry Image Registration Newton

Open paper AskPDF arXiv

Autonomous underwater navigation typically relies on complex and expensive multi-modal sensor suites designed to prioritize global Simultaneous Localization and Mapping (SLAM) accuracy. However, local reactive behaviors such as coarse navigation and obstacle avoidance require only local consistency---a capability that should be feasible using only a Forward-Looking Sonar (FLS), yet remains largely unaddressed, leaving a critical gap in FLS-only local SLAM. Moreover, existing acoustic SLAM frameworks predominantly rely on sparse feature extraction methods that discard substantial portions of the already information-sparse acoustic returns. To overcome these limitations, this work introduces a dense direct registration approach that aligns full acoustic intensity scans to a recursively updated local map. Real-time execution is achieved via an Inverse Compositional Gauss-Newton optimization strategy that minimizes computational overhead. Experimental evaluations show that this dense method yields significant improvements on translation error compared to sparse keypoint baselines, maintaining stable sub-meter tracking precision over wide displacement gaps. Moreover, this approach delivers odometry performance comparable to multi-sensor fusion pipelines (FLS, DVL, and IMU), bypassing expensive payload dependencies in feature-rich environments. We validate real-world applicability through AUV field trials, running the full local SLAM approach onboard an embedded, resource-constrained computer.

Abstract

Explore similar work

SOL-SLAM: Inverse Compositional Gauss-Newton Direct Registration for Fast Sonar-Only Local SLAM

SLAM in Low-Light Environments: Project Report

SLAMSqueezeBench: Comparing SLAM Systems under Resource Constraints