SaferPath: Hierarchical Visual Navigation with Learned Guidance and Safety-Constrained Control

TL;DR

SaferPath is a hierarchical visual navigation framework that leverages learned guidance from existing end-to-end models and refines it through a safety-constrained optimization-control module, enabling safe navigation in challenging real-world settings.

Abstract

Visual navigation is a core capability for mobile robots, yet end-to-end learning-based methods often struggle with generalization and safety in unseen, cluttered, or narrow environments. These limitations are especially pronounced in dense indoor settings, where collisions are likely and end-to-end models frequently fail. To address this, we propose SaferPath, a hierarchical visual navigation framework that leverages learned guidance from existing end-to-end models and refines it through a safety-constrained optimization-control module. SaferPath transforms visual observations into a traversable-area map and refines guidance trajectories using Model Predictive Stein Variational Evolution Strategy (MP-SVES), efficiently generating safe trajectories in only a few iterations. The refined trajectories are tracked by an MPC controller, ensuring robust navigation in complex environments. Extensive experiments in scenarios with unseen obstacles, dense unstructured spaces, and narrow corridors demonstrate that SaferPath consistently improves success rates and reduces collisions, outperforming representative baselines such as ViNT and NoMaD, and enabling safe navigation in challenging real-world settings.

Figures (7)

• Figure 1: SaferPath is a hierarchical navigation framework that uses a end-to-end visual navigation model for guidance and integrates safety-constrained control module, significantly improving obstacle avoidance and narrow-passage traversal for safe navigation in challenging environments.
• Figure 2: SaferPath Pipeline overview. SaferPath can be integrated with various end-to-end visual navigation models. The Traversability Score Mapper processes RGB observations to generate a score map $\mathcal{S}$, which, together with the learned guidance trajectory, is passed to the MP-SVES module. If optimization fails, the Emergency Indicator stops the robot and rotates it until a traversable trajectory is found; otherwise, the MPC controller executes trajectory tracking.
• Figure 3: Safety constrained trajectories under different safety thresholds, with the robot maintaining varying safe distances from obstacles by adjusting threshold $\delta$
• Figure 4: The figure illustrates the scene from Robust Navigation under Unseen Obstacles experiment. The scene above (Scene A) introduces obstacles along the sides of the road in the original environment, while the scene below (Scene B) includes all the obstacles from Scene A and adds additional obstacles in the middle of the road. The figures display some examples of successful trajectories of our method and examples of collision scenarios of the baseline methods.
• Figure 5: In Exploration in Dense Unstructured Environments experiment, the trajectory of SaferPath (Ours, NoMaD) (green) effectively avoids dense obstacles and reaches the free-space region. In contrast, the NoMaD trajectory (red) is more chaotic and results in significantly more collisions.