- The paper introduces a unified 3D traversability representation that integrates semantic guidance and hierarchical planning for effective cross-floor exploration.
- It demonstrates robust simulation performance on HM3D and MP3D, achieving state-of-the-art success and SPL improvements over existing baselines.
- Real-world deployment on a quadruped robot confirms its practical efficiency in navigating complex environments with stairs, narrow passages, and dynamic obstacles.
TravExplorer: Cross-Floor Embodied Exploration via Traversability-Aware 3D Planning
Problem Setting and Motivation
Zero-shot Object Navigation (ZSON) in unseen environments leverages vision-LLMs (VLMs) or LLMs to allow robots to perform open-vocabulary tasks such as searching for arbitrary user-specified objects. However, prevailing approaches remain encumbered by planarity and single-floor assumptions, which significantly limits real-world deployability. Real buildings feature cross-floor transitions, vertically overlapping rooms, and traversable structures (e.g., stairs, landings), whose representations in robot memory, as well as frontier extraction and planning, remain inadequately explored.
TravExplorer presents a paradigm shift by integrating traversability-aware volumetric mapping, lightweight yet robust semantic guidance, and a hierarchical planning architecture. The key innovation is a unified 3D traversability representation of the environment for quadruped robots, supporting efficient, physically feasible, and semantically guided exploration in both single and cross-floor scenarios, without prior maps or human interventions.
System Architecture
TravExplorer comprises three principal modules:
- Traversability-Aware Volumetric Mapping: RGB-D and odometry streams are integrated frame-wise to update a voxel-based occupancy map with explicit traversability and ESDF layers. Instead of modeling only free/occupied cells, TravExplorer distinguishes robot-reachable support surfaces (e.g., floors, stairs, landings) from true obstacles and traverses only within the former.
- Semantic Guidance: Lightweight probabilistic instance accumulation and image-text matching operate as a semantic map layer. Instead of full-scale LLM/VLM queries at each planning step, candidate objects and landmarks are extracted and fused over time, dramatically reducing semantic inference latency while increasing robustness to occlusions and partial observations.
- Hierarchical Exploration Planning: The planner operates over semantic and traversable frontiers, stair landmarks, and object hypotheses. A TSP-based tour selector, foothold-guided 3D graph search, and local trajectory optimizer convert the spatial-semantic reasoning output into real-world executable low-level actions.
Technical Innovations
3D Traversable Frontier Representation
Maps are updated in real-time with both occupancy and traversability. Adaptive fusion of geometric (2D ray casting on the support plane) and semantic (open-vocabulary segmentation) sources switches mode based on recent perceptual evidence, enabling robust detection of stairs and complex non-planar structures.
Frontiers are defined strictly on connected, reachable support surfaces with sufficient headroom. Clusters are efficiently updated incrementally and not globally, supporting real-time responsiveness. An FOV-aware active perception strategy (e.g., active tilting of the camera on approach to stairs or near frontiers) resolves local occlusions and blind spots.
Probabilistic Semantic Memory and Spatial Value Maps
Single-frame open-vocabulary detections tend to be noisy and unstable in exploration. TravExplorer maintains persistent instance maps, fusing temporally consistent detections and updating their belief via weighted point cloud integration and negative evidence (when expected but undetected). A second value map integrates fast image-text relevance scores (BLIP-2) projected onto reachable voxels, which efficiently guides exploration to regions evidentially associated with the current search target or key landmarks (e.g., stairs).
Hierarchical, Physically Feasible Planning
The planner hierarchically selects goals among robust semantic categories and connected traversable frontiers, formulating goal selection as an open asymmetric TSP over the current connectivity graph. Physical execution uses 3D path search restricted to traversable support surfaces, foothold-aware graph expansion, and local trajectory optimization using constrained B-splines. This ensures continuous, feasible, and dynamically smooth locomotion across both flat and vertically connected environments.
Execution is validated in a closed-loop: when execution deviates from planned path (e.g., due to newly encountered obstacles or navigation mismatches), the map is updated incrementally and the plan is revised.
Experimental Results
Simulation
TravExplorer was evaluated on HM3D (2000 episodes, 20 scenes) and MP3D (2195 episodes, 11 scenes) with standard SR (Success Rate) and SPL (Success weighted by Path Length) metrics. Compared to single-floor zero-shot (e.g., CoW, ApexNav, VLFM) and multi-floor ObjectNav baselines (MFNP, ASCENT), TravExplorer achieves state-of-the-art performance:
- HM3D: 70.0% SR, 37.2% SPL
- MP3D: 48.8% SR, 21.1% SPL
Notably, compared to ASCENT—the best cross-floor baseline—TravExplorer achieves gains of +4.6% SR and +3.7% SPL on HM3D, and +4.3% SR and +5.6% SPL on MP3D. The improvement is even more significant on multi-floor episodes, underscoring the efficacy of unified 3D traversability representations over prior floor-wise or 2D-centric approaches.
Failure Analysis
Dominant failure modes include false positives in semantic grounding (22% of cases), especially in cross-floor episodes, and step-budget exceedance for distant or multi-floor targets. Very few failures are due to exhausted frontiers, supporting the reliability of the proposed traversability-aware frontier extraction.
Real-World Deployment
TravExplorer was deployed on a Unitree Go2 quadruped with an RGB-D camera and LiDAR, tested across 50 trials in four real indoor environments (multi-floor villa, complex flat, cluttered loft, spacious office). The system achieved a 64.0% success rate without any prior maps or intervention, demonstrating consistent performance in real cross-floor scenarios. It reliably handled object search across stairs, selected semantically valuable frontiers, navigated narrow passages, and dynamically avoided moving obstacles.
Implications and Future Directions
TravExplorer delineates a pivotal advancement in cross-floor embodied exploration, bridging the gap between language-driven semantic target search and practical traversable geometry reasoning. By maintaining a unified, physically meaningful map and employing robust, low-latency semantic guidance, the system outperforms alternatives particularly in complex 3D and multi-floor settings.
This framework is directly extensible to broader classes of robots—including wheeled and tracked platforms—provided their traversability envelope is modeled appropriately. The approach also highlights the growing demand for fast, robust, and memory-efficient open-vocabulary semantic segmentation models for embodied agents.
Future work could expand TravExplorer's agentic autonomy by integrating more advanced high-level reasoning over unobserved regions, online learning of navigation policies tailored to unseen environments, and continuous improvement of semantic guidance via foundation model pretraining specific to robot deployment domains.
Conclusion
TravExplorer provides a comprehensive solution for zero-shot, cross-floor embodied object navigation in unknown indoor environments. Its integration of traversability-aware 3D mapping, lightweight semantic guidance, and hierarchical planning yields robust, generalizable, and physically executable exploration exceeding previous benchmarks in both simulation and real-world settings. The results assert the superiority of unified 3D semantic-geometric representations over planar or floor-wise abstractions for real-world embodied exploration (2605.19958).