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

Can sequence verification and a robust pose-graph back-end make sonar-based place recognition reliable enough for loop closure in self-similar environments, without map collapse?

Sonar place recognition is ambiguous: corridors, junctions, or rows of pillars can produce nearly identical echo trains, causing false loop closures that collapse topological maps; existing BatSLAM used single local-view comparisons and could inject wrong matches with full confidence.

Earlier BatSLAM combined biomimetic binaural sonar with RatSLAM pose cells and experience map, comparing complete cochleogram templates as local views. It recognized places but lacked sequence-level validation and a principled way to reject or remove false loop closures. More general visual/lidar SLAM methods such as SeqSLAM, switchable constraints, and dynamic covariance scaling exist, but their transfer to ambiguous sonar views had not been tested.

BatSLAM 2.0 replaces pose cells with a GTSAM factor graph solved by iSAM2. The acoustic front-end builds a local view from two images per ear: an energy image and a spectral-shape image that encodes direction via ear/emitter filtering. Place matches are searched by correlation with range shift. A sequence verifier maintains competing recognition hypotheses, accumulates evidence from matches, checks geometric consistency, and commits only when a hypothesis is long, strong, unambiguous, and plausible under current pose uncertainty. Committed links enter the pose graph as weak constraints grouped by hypothesis; link management can withdraw groups using a length rule, residual check, and periodic GNC audits.

Simulation: with the design sonar (10 m range, -116 dB noise) and calibrated or roughly calibrated odometry, all 12 runs produced consistent maps: 0 wrong commits out of 457, link precision 0.994–1.000, 0.79–0.84 of true same-direction revisits linked, and absolute trajectory error 0.47–1.91 m unaligned / 0.18–0.53 m aligned, versus 4.2–19.0 m odometry. Single-view top-1 recognition reached 0.89 for the final descriptor in the city; adding spectral-shape raised indoor top-1 from 0.88 to 0.91. Ablations: without sequence verification, 215 wrong commits occurred and 10 survived link management, mean ATE rose to 4.85 m; removing link management too caused map collapse. Template capture region was about ±0.15 m laterally and ±25° in heading. Real-world beamformed eRTIS sonar: across 30 runs, no wrong hypotheses or collapses; with a refitted evidence curve, trajectory error fell from 3.62 (0.46) m to 0.67 (0.13) m, and 0.64 (0.10) of revisits were linked; per-pulse compute was 4.8 ms front-end plus 6.2 ms rest vs 141 ms between pulses.

Most evaluation is in simulation; the simulator omits weak higher-order reflections. Real-world validation used only one beamformed 3D sonar in one hall, requiring per-sensor recalibration of the evidence curve. The system cannot recognize places revisited only in the opposite direction. It assumes odometry errors are predominantly random, so systematic heading bias degrades metric accuracy (though topology remains correct). The GNC audit never changed a final map in these experiments, so its benefit is unproven. Template capture is narrow (±0.15 m laterally, ±25° heading), which limits robustness to pose variation.