DR. Nav: Semantic-Geometric Representations for Proactive Dead-End Recovery and Navigation

TL;DR

DR. Nav tackles the challenge of safe autonomous navigation in unstructured environments by predicting dead-ends and planning recoveries before entrapment. It integrates RGB–LiDAR fusion with cross-attention into a continuous semantic cost map that encodes per-cell dead-end risk and recovery points, updated via Bayesian filtering. A short-horizon planner uses an Expected Dead-End Exposure term to bias paths away from risky zones, while a DWA controller executes safe trajectories and leverages recovery options when needed. Empirical results on dense indoor/outdoor scenarios show substantial gains in dead-end detection accuracy and path efficiency compared with reactive planners and image-only detectors, highlighting the practical impact of proactive, semantics-aware navigation. The approach enables safer, more reliable navigation in challenging environments and points toward real-time deployment and richer recoverability modeling in future work.

Abstract

We present DR. Nav (Dead-End Recovery-aware Navigation), a novel approach to autonomous navigation in scenarios where dead-end detection and recovery are critical, particularly in unstructured environments where robots must handle corners, vegetation occlusions, and blocked junctions. DR. Nav introduces a proactive strategy for navigation in unmapped environments without prior assumptions. Our method unifies dead-end prediction and recovery by generating a single, continuous, real-time semantic cost map. Specifically, DR. Nav leverages cross-modal RGB-LiDAR fusion with attention-based filtering to estimate per-cell dead-end likelihoods and recovery points, which are continuously updated through Bayesian inference to enhance robustness. Unlike prior mapping methods that only encode traversability, DR. Nav explicitly incorporates recovery-aware risk into the navigation cost map, enabling robots to anticipate unsafe regions and plan safer alternative trajectories. We evaluate DR. Nav across multiple dense indoor and outdoor scenarios and demonstrate an increase of 83.33% in accuracy in detection, a 52.4% reduction in time-to-goal (path efficiency), compared to state-of-the-art planners such as DWA, MPPI, and Nav2 DWB. Furthermore, the dead-end classifier functions

Figures (4)

• Figure 1: The continuous cost map encodes environment semantics at the cell level, where green regions represent navigable areas and red regions correspond to predicted dead-ends or risky zones and the recovery point as blue cylinders . Unlike a binary occupancy grid, this representation continuously updates as new RGB–LiDAR observations arrive, allowing the robot to reason about both safety and recoverability. In this example, the robot (yellow marker) progresses forward as long as at least one side of its camera view remains open. When the surrounding views become fully blocked, the system detects the impending dead-end (red region ahead) and proactively backtracks to a pre-identified recovery point (blue marker). This illustrates how the continuous cost map not only predicts unsafe regions in advance but also guides the robot toward safe alternatives in real time.).
• Figure 2: Our method uses a multi-modal sensor fusion architecture for costmap computations and autonomous navigation. Our proposed model processes RGB images through an EfficientNet encoder and LiDAR point clouds through a PointNet encoder, applying dual-stage spatial attention mechanisms to enhance feature representations. We perform cross-modal fusion to integrate the enhanced multi-modal features, which are then processed through an integration MLP to generate path classification, dead-end detection, and cost map outputs for autonomous navigation tasks.
• Figure 3: Qualitative results of dead-end detection and risk estimation
• Figure 4: Qualitative results. These qualitative results show that DR. Nav introduces proactive divergent trajectories compared to current planners that fail to account for the dead-ends, become trapped, and then backtrack.