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

Technical Summary: EffiNav

Problem Statement
Object Goal Navigation (ObjNav) requires an autonomous agent to locate a specific target object in an unknown environment without prior map knowledge. While existing approaches have achieved success in reaching targets, they often struggle with navigation efficiency, leading to excessive exploration, redundant back-and-forth motion, and poor generalization. Current solutions face a trade-off:

• End-to-end foundation models offer versatility but require massive computational resources and large-scale training data, often failing to generalize well to diverse object categories.
• Learning-based modular architectures (e.g., diffusion models) are sensitive to training data quality and struggle with long-horizon tasks and environmental diversity.
• Training-free frameworks leveraging Vision-Language Models (VLMs) offer robustness and zero-shot generalization but often suffer from inefficiencies, such as getting trapped in corners or revisiting explored areas due to a lack of historical memory and global spatial reasoning.

Methodology
The authors propose EffiNav, a training-free, depth-aware framework that fuses visual observations with top-down spatial reasoning to enable efficient navigation. The system relies on an open-source VLM (Qwen2.5VL-7B) as its reasoning core and utilizes an RGB-D sensor and odometry for localization. The framework operates through the following modules:

1. Perception and Intelligent Guidance:

   • Ego Region Selecting (Module a): The system generates candidate exploration masks based on depth thresholds (e.g., areas >5m away). These regions are labeled and presented to the VLM via a Vision Question Answering (VQA) prompt (e.g., "To find the bed, which area is best?") to select the most promising local direction.
   • Global-wise Checking (Module b): The selected region is projected onto a gradually constructed top-down map. This map is fed back to the VLM with a consistency check prompt (e.g., "Is the colored area a good option?"). This step validates the spatial reasonableness of the choice from a global perspective.
   • History-Aware Pruning: To prevent redundant exploration, candidates lying within a predefined distance of already explored areas are pruned. If all candidates are rejected, the system falls back to the nearest frontier point to ensure continued exploration.

2. Destination Confirmation and Path Planning:

   • Destination Confirmation (Module c): The nearest valid candidate from the top-down projection is selected as the immediate goal. If unreachable, the destination mask is iteratively dilated.
   • Path Planning (Module d): An A* search algorithm computes a collision-free trajectory to the confirmed goal. The map is pre-processed by dilating obstacle regions to account for reconstruction imperfections. Discrete low-level commands (e.g., "turn left," "move forward") are generated based on the robot's state relative to the path.

3. Map Construction and Termination:

   • The environment map is built incrementally using depth and odometry, starting with a 360° rotation.
   • Navigation terminates when an object detector identifies the target in the ego-observation. The object is segmented, projected onto the top-down map, and the robot navigates directly to this region without further VLM guidance.

Key Contributions

• Novel Framework: Introduction of EffiNav, a training-free framework that integrates depth-aware perception with VLM reasoning to balance local exploration and global consistency.
• Efficiency Metric: Proposal of Normalized Efficiency on Successes (EoS), a new metric defined as $SPL / SR$, to quantify navigation efficiency specifically within successful episodes, offering a finer-grained assessment than SR or SPL alone.
• Extensibility: Demonstration of the framework's adaptability by extending it to Memory-augmented ObjNav on the GOAT-BENCH dataset, where it incorporates a memory module to track previously found objects.
• Real-World Validation: Successful deployment on a physical Unitree Go2 robot in a lab environment, validating the framework beyond simulation.

Experimental Results
EffiNav was evaluated on two standard benchmarks: Habitat Matterport 3D (HM3D) and Open-Vocabulary Object Goal Navigation (OVON), as well as the GOAT-BENCH memory task.

• Performance: EffiNav achieved a Success Rate (SR) of 55.0% and Success weighted by Path Length (SPL) of 36.7 on HM3D, and 63.0% SR and 43.7% SPL on OVON.
• Efficiency: The framework achieved the highest EoS scores on both datasets (66.8 on HM3D and 69.1 on OVON), outperforming nearly all baselines, including large-scale trained foundation models like UniNav.
• Comparison: While UniNav achieved higher SR on HM3D, it required significant training resources (483k samples, 40 GPUs) and performed poorly on OVON. EffiNav, requiring no training, demonstrated superior generalization and efficiency, ranking second on HM3D and first on OVON.
• Stability: Analysis showed EffiNav maintains consistent EoS across diverse scenes (Coefficient of Variation ~0.085 on HM3D), indicating robustness against environmental variations.
• Real-World: In physical robot experiments, EffiNav outperformed a nearest-frontier baseline in both SR (0.80 vs. 0.60) and SPL (0.65 vs. 0.43).

Significance and Claims
The paper claims that EffiNav addresses the critical inefficiency issues in training-free navigation by combining VLM reasoning with depth-aware global verification. The authors assert that the framework offers a "balanced and generalizable" approach that rivals or exceeds state-of-the-art trained models without the associated costs of data collection and training. The work highlights that intelligent exploration, driven by the fusion of egocentric visual cues and top-down spatial reasoning, is key to efficient ObjNav. The successful extension to memory-augmented tasks and real-world deployment underscores the framework's practical applicability and potential for broader robotic navigation tasks. The authors note that as VLM capabilities evolve, the potential of this framework is expected to grow further.