Research map / Vision-language and multimodal / Multimodal LLMs

GA-VLN: Geometry-Aware BEV Representation for Efficient Vision-Language Navigation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GA-VLN, which feeds a compact geometry-aware BEV representation (depth-projected visual features fused with VGGT 3D-foundation-model features) into an MLLM navigator, achieving state-of-the-art continuous VLN results on R2R-CE, RxR-CE and NavRAG-CE without DAgger or VQA co-training.

Problem

Image-based MLLM navigation agents process dense RGB video patch tokens, which are token-redundant and lack explicit 3D spatial structure, causing high computational cost and weak spatial reasoning in continuous vision-language navigation.

Approach

Lift RGB-D patch features into 3D using depth and camera pose, aggregate them into an agent-centric BEV grid with mean pooling over non-empty cells, and fuse in features from a frozen 3D foundation model (VGGT) projected into the same BEV space; the BEV tokens replace dense historical frame tokens in an MLLM (LLaVA-Video-7B) action predictor.

Key results

• On R2R-CE val_unseen, GA-VLN reaches NE 4.80, OSR 67.6, SR 61.0, SPL 55.2, exceeding InternVLA-N1 (SR 58.2, SPL 54.0) without DAgger.
• Ablation on R2R-CE: GA-BEV reduces visual tokens from 4003 to 514 (with VGGT) while SR rises from 46.49 to 53.56 in the no-SRDF setting.
• On RxR-CE val_unseen, GA-VLN achieves SR 55.4 and NE 5.88, above Uni-NaVid and NaVILA on SR.

Surprising findings

• Naively concatenating depth-encoded tokens with RGB tokens doubles token count and degrades all metrics (SR 46.49 to 38.61), whereas BEV projection improves them.

Limitations

• Zero-shot real-world transfer shows occasional paths hugging walls and imprecise stopping due to lack of obstacle avoidance.
• Depends on depth input (RGB-D) which the monocular benchmark setting adds beyond pure RGB baselines.
• Reported NavRAG-CE gains come from additional fine-tuning, and including NavRAG-CE in pretraining hurt R2R-CE/RxR-CE generalization.

Experimental setup

Datasets: R2R-CE RxR-CE EnvDrop ScaleVLN SRDF NavRAG-CE

Benchmarks: R2R-CE val_unseen RxR-CE val_unseen NavRAG-CE val_unseen

Metrics: NE OSR SR SPL

Baselines: NaVILA StreamVLN InternVLA-N1 Uni-NaVid NaVid Aux-Think MonoDream Dynam3D g3D-LF MapNav VLN-3DFF

Builds on: NaVILA NaVid VGGT LLaVA-Video DUET

Closest papers by meaning

• Dynam3D: Dynamic Layered 3D Tokens Empower VLM for Vision-and-Language Navigation (NeurIPS 2025 · similarity 0.90)
• g3D-LF: Generalizable 3D-Language Feature Fields for Embodied Tasks (CVPR 2025 · similarity 0.89)
• Volumetric Environment Representation for Vision-Language Navigation (CVPR 2024 · similarity 0.89)
• Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.88)
• Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method (CVPR 2025 · similarity 0.88)
• Lookahead Exploration with Neural Radiance Representation for Continuous Vision-Language Navigation (CVPR 2024 · similarity 0.87)