AquaBEV-Nav: Learned BEV Occupancy for Underwater Navigation and Exploration

Trung Tien Dong; Zhenqi Wu; Sahasra Kondapalli; Jiayi Wu; Yi Sheng; Xiaomin Lin

cs.ROSep 26, 2026

AquaBEV-Nav: Learned BEV Occupancy for Underwater Navigation and Exploration

Authors: Trung Tien Dong, Zhenqi Wu, Sahasra Kondapalli, Jiayi Wu, Yi Sheng, Xiaomin Lin

Organizations: 1ERA Lab, University of South Florida, Florida, USA · University of Maryland- College Park, College Park, Maryland · 2YES Lab, University of South Florida, Florida, USA

Abstract

Safe underwater exploration requires a robot to understand where surrounding structures are located and which regions are available for motion. Existing vision-based underwater exploration systems commonly obtain this information indirectly by estimating monocular depth, unprojecting the geometry into 3D space, and accumulating it into a 2D bird's-eye-view occupancy map. This reliance on intermediate depth estimation is particularly problematic underwater, where scattering and wavelength-dependent attenuation degrade visual cues and limit the reliability of monocular depth estimates. We introduce AquaBEV-Nav, an underwater exploration framework that bypasses explicit monocular depth estimation through direct bird's-eye-view occupancy prediction. Built upon the CORAL hierarchical exploration framework, AquaBEV-Nav replaces its depth-based perception front end with AquaBEV. Given a single RGB frame, AquaBEV maps visual features into a learned polar representation, performs causal reasoning along the range dimension, and reconstructs local Cartesian occupancy without relying on intermediate depth prediction. The resulting occupancy map is accumulated into CORAL's persistent spatial memory, providing spatial context for VLM-based high-level planning and collision constraints for dynamics-aware local trajectory generation. Across ten simulated reef environments and six occupancy backbones evaluated under a single protocol, AquaBEV-Nav reaches 37.48 structure IoU and 53.2 target IoU, 88.95% closed-loop coverage with zero collisions.

Figures & tables

Fig. 1: AquaBEV-Nav. CORAL constructs occupancy through monocular depth estimation, geometric reconstruction, and rasterization (top). AquaBEV-Nav replaces this perception pipeline with direct monocular BEV occupancy prediction (bottom).

Fig. 2: AquaBEV-Nav architecture. AquaBEV-Nav predicts local structure, target, and observability fields from a single RGB image. These fields are transformed into the world frame and accumulated in navigation system persistent spatial memory.

	Occupancy			Structure $\rightarrow$ grid			Target $\rightarrow$ seg. map
Arm	mIoU $\uparrow$	Prec. % $\uparrow$	Recall % $\uparrow$	mIoU $\uparrow$	Prec. % $\uparrow$	Recall % $\uparrow$	mIoU $\uparrow$	Prec. % $\uparrow$	Recall % $\uparrow$
CORAL (DAv2) [ 5 ]	—	—	—	10.0	10.6	65.2	11.0	18.3	21.5
TPVFormer [ 24 ]	14.8	76.1	15.6	11.0	22.2	18.0	38.6	45.8	71.0
SurroundOcc [ 25 ]	16.7	75.3	17.7	16.0	31.7	24.5	38.9	47.4	68.3
MonoScene [ 33 ]	24.1	80.4	25.6	21.5	26.8	52.4	39.7	43.5	82.0
GaussianFormer-fix [ 7 ]	23.5	80.0	24.9	27.6	31.0	71.5	41.8	47.0	78.8

TABLE I: Planner Facing BEV Prediction Performance. Comparison of occupancy, structure, and target prediction under a unified training protocol. All learned models are initialized from their respective real world checkpoints, anchored to frozen occupancy teachers, and trained for 25 epochs under the same schedule.( BEST / SECOND BEST )

	Structure				Target
Arm	max IoU $\uparrow$	max P $\uparrow$	mean IoU $\uparrow$	mean P $\uparrow$	max IoU $\uparrow$	max P $\uparrow$	mean IoU $\uparrow$	mean P $\uparrow$
CORAL (DAv2) [ 5 ]	0 8.3	0 8.3	21.5	25.7	25.0	28.7	0 8.5	20.7
TPVFormer [ 24 ]	21.9	25.1	0 7.9	47.7	47.5	48.8	47.8	67.7
SurroundOcc [ 25 ]	27.4	32.0	11.8	71.7	47.6	49.1	46.2	71.8
MonoScene [ 33 ]	22.5	23.7	31.3	56.7	46.0	46.6	53.2	63.0
GaussianFormer-fix [ 7 ]	27.1	28.0	44.7	55.4	52.2	53.8	53.8	67.5

TABLE II: Accumulated planner map performance. Comparison of structure and target maps after trajectory level fusion. ( BEST / SECOND BEST )

Fig. 3: Simulation perception. Qualitative comparison of local BEV predictions across representative environments.

Method	Structure		Target
Method	mIoU $\uparrow$	Prec. % $\uparrow$	mIoU $\uparrow$	Prec. % $\uparrow$
AquaBEV	29.47 $\pm$ 0.31	39.0 $\pm$ 1.8	46.39 $\pm$ 0.50	51.9 $\pm$ 1.5
+ ray PE	29.28 $\pm$ 0.76	39.1 $\pm$ 2.0	45.94 $\pm$ 0.50	51.2 $\pm$ 1.3
+ ray consist.	29.24 $\pm$ 0.55	39.3 $\pm$ 2.7	45.91 $\pm$ 0.25	50.5 $\pm$ 0.8
+ ray PE & consist.	29.30 $\pm$ 0.07	38.5 $\pm$ 0.3	45.81 $\pm$ 0.91	50.8 $\pm$ 2.3
+ range CE	30.35 $\pm$ 0.41	39.8 $\pm$ 1.4	47.10 $\pm$ 0.11	52.1 $\pm$ 0.9

TABLE III: Architecture ablation for planner field prediction. Comparison of structural changes applied to the AquaBEV backbone.

Method	Structure		Target
Method	mIoU $\uparrow$	Prec. % $\uparrow$	mIoU $\uparrow$	Prec. % $\uparrow$
AquaBEV	29.47 $\pm$ 0.31	39.8 $\pm$ 1.4	46.39 $\pm$ 0.50	51.9 $\pm$ 1.5
+ $\mathcal{L}_{\mathrm{rad}}$	36.00 $\pm$ 0.71	49.5 $\pm$ 1.7	46.32 $\pm$ 0.65	50.4 $\pm$ 1.5
+ $\mathcal{L}_{\mathrm{sfp}}$	35.96 $\pm$ 0.76	49.9 $\pm$ 2.9	48.54 $\pm$ 0.38	53.6 $\pm$ 1.5
+ $\mathcal{L}_{\mathrm{ot}}$	36.26 $\pm$ 0.56	51.7 $\pm$ 4.1	52.74 $\pm$ 0.20	59.6 $\pm$ 1.0
+ distillation	37.39 $\pm$ 0.62	54.6 $\pm$ 0.4	51.74 $\pm$ 0.29	59.1 $\pm$ 0.2

TABLE IV: Planner aware objective ablation. Effect of the proposed field specific training objectives on structure and target prediction.

Fig. 4: Closed loop navigation by reef topology. Executed trajectories produced by different perception front ends under the same CORAL navigation stack.

Front end	cov.% $\uparrow$	time s $\downarrow$	coll. $\downarrow$	VLM $\downarrow$	dev. $\downarrow$
DREAM [ 4 ]	80.00	1513.2	0 9	1261	15
CORAL (DA2)	59.58	620	0 0	0 519	0 3
+ TPVFormer [ 24 ]	68.33	734	0	191	19
+ GaussianFormer [ 7 ]	43.21	612	0	394	19
+ SurroundOcc [ 25 ]	49.17	669	0	188	29
+ AquaBEV-Nav (ours)	88.95	734	0	353	5

TABLE V: Closed-loop simulation navigation comparison across ten reef environments.

Fig. 5: Pool perception. Representative front-view RGB observations and corresponding local BEV predictions from AquaBEV-Nav and the depth-based CORAL front end.

Explore similar work

Sep 3, 2026

AquaBEV: Monocular Underwater BEV Occupancy with 3D Sonar Supervision

Autonomous underwater robots are widely used for exploration, monitoring, and inspection, where safe navigation depends on understanding the surrounding free and occupied space. Bird's eye view (BEV) occupancy provides such a representation, but predicting it from a single underwater RGB image is difficult due to limited, unreliable geometric cues from appearance alone. 3D imaging sonar offers complementary geometric measurements to supervise this task. We introduce AquaBEV, a monocular underwater occupancy model that predicts local BEV occupancy from a single RGB image, using paired 3D imaging sonar as geometric supervision during training. AquaBEV maps visual features into a calibration free polar representation and applies causal decoding along the range dimension before reconstructing the prediction in Cartesian BEV coordinates. A controlled underwater occupancy benchmark was established, adapting representative occupancy methods to the same RGB to sonar task under a unified protocol. AquaBEV achieves 31.4 Visible IoU and 38.6 Observed IoU, 4.0% and 4.3% relative improvements over the strongest transferred baseline.

Trung Tien Dong, Shengji Jin, Chen Chen +2

ERA Lab, University of South Florida, Tampa, Florida, USA. · YES Lab, University of South Florida, Tampa, Florida, USA. · University of Central Florida, Orlando, Florida, USA.

3D Occupancy Prediction Underwater Robotics Underwater Imaging

Ask PDF

Mar 16, 2026

CORAL: COntextual Reasoning And Local Planning in A Hierarchical VLM Framework for Underwater Monitoring

Oyster reefs are critical ecosystem species that sustain biodiversity, filter water, and protect coastlines, yet they continue to decline globally. Restoring these ecosystems requires regular underwater monitoring to assess reef health, a task that remains costly, hazardous, and limited when performed by human divers. Autonomous underwater vehicles (AUVs) offer a promising alternative, but existing AUVs rely on geometry-based navigation that cannot interpret scene semantics. Recent vision-language models (VLMs) enable semantic reasoning for intelligent exploration, but existing VLM-driven systems adopt an end-to-end paradigm, introducing three key limitations. First, these systems require the VLM to generate every navigation decision, forcing frequent waits for inference. Second, VLMs cannot model robot dynamics, causing collisions in cluttered environments. Third, limited self-correction allows small deviations to accumulate into large path errors. To address these limitations, we propose CORAL, a framework that decouples high-level semantic reasoning from low-level reactive control. The VLM provides high-level exploration guidance by selecting waypoints, while a dynamics-based planner handles low-level collision-free execution. A geometric verification module validates waypoints and triggers replanning when needed. Compared with the previous state-of-the-art, CORAL improves coverage by 14.28% percentage points, or 17.85% relatively, reduces collisions by 100%, and requires 57% fewer VLM calls.

Zhenqi Wu, Yuanjie Lu, Xuesu Xiao +1

University of South Florida · George Mason University

Autonomous Underwater Navigation Vision-Language Navigation VLMs for Robotics Large Language Model-Based Robot Planning

Ask PDF

Aug 28, 2026

CAVE-NAV: VLM-Based Autonomous 3D Navigation in Underwater Cave Environments

Autonomous navigation in underwater cave environments is essential for search-and-rescue operations, scientific exploration, and emergency egress. Traditional navigation systems commonly depend on dense visual features for localization and mapping. In underwater caves, however, visual degradation can undermine feature-based localization, sonar-based mapping may yield overly conservative obstacle representations, and communication constraints preclude real-time human guidance. To address these limitations, we propose an autonomous underwater cave navigation framework that leverages a vision-language model (VLM) with Chain-of-Thought (CoT) reasoning to infer navigable directions from environmental cues, including light intensity gradients, passage morphology, and geometric complexity, captured through multimodal inputs comprising RGB imagery, depth maps, and sonar-based vertical-clearance measurements, thereby supporting safe 3D navigation through confined cave passages. High-fidelity simulations across multiple cave topologies demonstrate that the proposed framework completes all evaluated end-to-end traversals without collisions while maintaining safe clearance from cave boundaries.

Zhenqi Wu, Yuanjie Lu, Yisheng Zhang +4

Embodied Robotics and Automation Lab, University of South Florida, Tampa, FL 33620, USA · Computer Science Department, George Mason University, Fairfax, VA 22032, USA · Department of Mechanical Engineering, University of Maryland, College Park, MD 20742, USA +1

Autonomous Underwater Navigation Vision-Language Navigation VLMs for Robotics

Ask PDF