EffiNav: Fusing Depth and Vision-Language for Efficient Object Goal Navigation

Abstract: To locate a target object while exploring the unknown environment is a fundamental capability for autonomous agents, with applications ranging from search-and-rescue to field robots. A simplified version of such task is Object Goal Navigation (ObjNav). In ObjNav, successful arrival at the target object provides a basic measure of performance; however, the efficiency of the navigation trajectory is equally important, as it indicates how intelligently the agent explores and how much time remains for subsequent tasks. In unknown environments, the key to efficient navigation lies in deciding where to explore next. While many prior works aim to address this core challenge and achieved promising performance in certain settings, recent training-based models and non-training frameworks still suffer from generalization and efficiency issues respectively, which in the worst cases can lead to excessive exploration of already-visited areas or redundant back-and-forth motion. We evaluate EffiNav on two widely used simulation benchmarks Habitat Matterport 3D (HM3D) and Open-Vocabulary Object goal Navigation (OVON), and further validate its effectiveness on physical robots in real-world settings. We conduct failure analysis on massive simulation episodes. With minimal modification, we also extend EffiNav to a memory-augmented ObjNav task on the GOAT-BENCH dataset, demonstrating its adaptability beyond standard ObjNav settings. Across two standard metrics--Success Rate (SR) and Success weighted by Path Length (SPL), EffiNav matches or outperforms recent baselines, reflecting its efficiency, robustness, and practical applicability. Recognizing the different emphases of the two datasets, the performances reveals this framework is more balanced and generalizable for efficient ObjNav.

• The paper introduces a training-free RGB-D navigation framework that combines depth-gated region selection, vision-language verification, history-aware pruning, and A* replanning to reduce redundant exploration.
• EffiNav achieves the highest EoS on HM3D and OVON—66.8 and 69.1—and leads OVON success rate with 63.0%, while trailing trained Uni-NaVid on HM3D success rate.
• The results show strong open-vocabulary and path-efficiency performance, but reliable depth sensing, 2D viewpoint limitations, semantic confusion, and unreported VLM inference costs constrain deployment.

Overview and motivation

EffiNav is a training-free framework for Object Goal Navigation (ObjNav) that combines depth-aware egocentric perception with vision-language reasoning to improve navigation efficiency. The authors, Yin and Ma of HKUST(GZ), position the work against two dominant paradigms: end-to-end navigation foundation models and learning-based modular architectures. Their central argument is that both suffer from complementary failure modes—foundation models generalize poorly across object categories despite massive training budgets, while training-free frontier-based methods waste trajectory length through redundant re-exploration and back-and-forth motion between distant frontiers.

The paper's efficiency claim is quantified through a new metric, Normalized Efficiency on Successes (EoS), defined as SPL divided by SR. Since successful episodes satisfy $L_i \geq L_i^*$, EoS reduces to the mean path ratio $L_i^*/L_i$ over successes only, decoupling trajectory efficiency from raw success rate. This decomposition is useful for diagnosing whether a method's low SPL stems from low success or from inefficient paths—a distinction that matters when comparing methods with very different SR profiles.

Framework design

EffiNav requires only an RGB-D sensor and odometry, using Qwen2.5-VL-7B as its decision core and MobileSAM for segmentation. The pipeline has four modules:

• Ego region selecting: depth-thresholded candidate regions (>5 m) are labeled in the egocentric view; a VQA prompt asks the VLM which numbered area is most promising for finding the target.
• Global-wise checking: the selected region is projected onto an incrementally built top-down map, recolored, and fed back to the same VLM for a binary spatial-reasonableness check. Rejected candidates fall back to the next-best option; if all are rejected, the nearest map-boundary frontier point serves as a fallback target—though the authors report this fallback is rarely invoked. A history-aware pruning mechanism removes candidates lying beyond a margin from already-explored areas, directly targeting the re-exploration pathology of prior VLM-based schemes.
• Destination confirmation and path planning: the nearest reachable projected candidate becomes the short-term goal, with iterative mask dilation for feasibility; A* planning on an obstacle-dilated map handles dynamic replanning as new obstacles appear.
• Finalization: Qwen2.5-VL detection plus MobileSAM segmentation projects the goal instance onto the top-down map; termination is pixel-distance-based with periodic goal refinement every five steps.

Notably, the authors state they invested little effort in prompt engineering—the prompts elicit only a number or a yes/no answer—and kept them fixed across all three benchmarks.

Benchmark results

The evaluation covers HM3D val (2,019 episodes, 20 scenes) and OVON unseen_val (3,000 episodes, 35 scenes), chosen deliberately because they stress different capabilities: HM3D has longer geodesic distances (mean 7.60 m vs. 5.81 m) and few object categories, while OVON features long-tailed object vocabulary but many goals within detection range of the start pose.

Method	Training cost	HM3D SR / SPL / EoS	OVON SR / SPL / EoS
PIRLNav	*	64.1 / 27.1 / 42.3	–
CogNav		72.5 / 26.2 / 36.1	–
VLFM		52.5 / 30.4 / 57.9	35.2 / 19.6 / 55.7
Uni-NaVid	••*	73.7 / 37.1 / 50.3	39.5 / 19.8 / 50.1
NavFom	••*	–	43.6 / 31.3 / 71.8
EffiNav	none	55.0 / 36.7 / 66.8	63.0 / 43.7 / 69.1

Three findings stand out. First, EffiNav achieves the highest EoS on both benchmarks (66.8 HM3D, 69.1 OVON), supporting the paper's core efficiency claim. Second, it ranks first on OVON overall (63.0 SR), substantially outperforming Uni-NaVid (39.5)—a model trained on 483,000 HM3D samples over roughly 1,400 GPU-hours—which degrades sharply under open-vocabulary goals. Third, on HM3D it trails Uni-NaVid by nearly 19 points of SR (55.0 vs. 73.7); the paper is candid that the trained foundation model retains the top spot where navigation difficulty dominates. Per-scene EoS is notably stable (coefficient of variation 0.083–0.113) even though per-scene SR fluctuates widely with scene geometry, suggesting the efficiency gains are robust rather than concentrated in easy scenes.

On GOAT-BENCH memory-augmented ObjNav, a minimal extension recording previously found object locations yields 52.3 SR and 48.5 SPL versus 49.6/29.4 for 3D-Mem—the SPL gap being the largest reported margin in the paper. The authors note their implementation performs full path planning from scratch, whereas 3D-Mem relies on Habitat planner teleportation, which strengthens but also complicates this comparison.

Real-world validation on a Unitree Go2 with an Azure Kinect in an office environment (5 episodes, goal "teddy bear on sofa") shows EffiNav at 0.80 SR / 0.65 SPL / 0.81 EoS against 0.60/0.43/0.72 for BFS nearest-frontier exploration. The sample size is small enough that these numbers should be read as qualitative evidence rather than statistical support.

Failure analysis

The authors analyze all failures—906 of 2,019 episodes on HM3D and 993 of 3,000 on OVON—categorizing them into constraint-map exhaustion, semantic confusion (e.g., lamp vs. table lamp; "shelf" and "kitchen cabinet" are hardest), category mismatch in detection, reconstruction artifacts, and timeout. The failure distributions differ instructively between benchmarks: constraint-map and timeout dominate HM3D (~45% constraint-map), while semantic understanding and category mismatch dominate OVON (~25% constraint-map). This mirrors the datasets' design emphases and indicates that EffiNav's residual errors on HM3D are exploration-capacity problems, not reasoning problems—an honest framing given its SR deficit there.

Limitations

The paper explicitly concedes three limitations. First, the pipeline reasons over 2D images (egocentric and top-down) because current training-free VLMs cannot reliably perform multi-view 3D correspondence, sacrificing spatial information by design. Second, the framework is critically depth-dependent: projection errors propagate directly into guidance quality, and the authors' attempt to substitute monocular estimated depth (Depth Anything 3) failed due to temporal scale inconsistency, leaving accurate RGB-D sensing a hard prerequisite. Third, finalization relies on 2D detection sensitive to viewpoint, contributing the skewed-view table-as-bed failures observed on HM3D. An additional unstated concern is inference cost: every decision step invokes one or more VLM queries, and the paper does not report wall-clock latency or query counts, which bear on real-time deployability.

Conclusion

EffiNav demonstrates that a carefully engineered training-free pipeline—depth-gated candidate generation, dual egocentric/global VQA verification, and history-aware pruning—can match or exceed heavily trained navigation foundation models on open-vocabulary ObjNav while achieving the best measured path efficiency among compared methods. Its principal trade-off is a clear SR deficit on navigation-heavy HM3D relative to Uni-NaVid, and its dependence on reliable depth sensing constrains immediate real-world transfer. The open questions the work leaves are concrete: whether VLM-based global verification can be extended to genuine 3D scene representations without training, and whether the efficiency advantage persists under larger real-world evaluation protocols than the five-episode validation reported here.