In crowded dynamic scenes, autonomous navigation must satisfy three key requirements: geometry safety, real-time performance, and interpretability. The mainstream pipelines typically consider only one or two requirements. Therefore, we propose SDERI, a hierarchical navigation framework that includes safe region subgoal distribution and exploration-replanning index (ERI). First, we bound the region of the subgoal distribution to a safety corridor that is consistent with global reachability. Then, an ERI is used to regulate the entropy of the subgoal distribution within the safe region and the replan timing of the current subgoal distribution. ERI is computed from minimal online cues, local occupancy ratio and normalized remaining distance. The physical simulation experiments demonstrate that our navigation method is effective in diverse dynamic scenarios. Meanwhile, it outperforms the baseline algorithms in terms of success rate and safety, especially in the case where the environment's  crowd density increases. Moreover, the ablation experiments confirm the necessity of each component in this framework.

Subgoal Distributions and Replanning Times under Different ERIs

Panels are ordered left→right, top→bottom with ERI = 0.817, 0.603, 0.375, 0.129. The background is a 100×100 local occupancy grid (the dark gray area is obstacles); the black lines are the global paths generated by RRT that are reachability-equivalent; the red dot marks the robot, and the translucent red disc is the near-field exclusion zone that is removed from sampling. The colored heatmap shows the subgoal distribution restricted to the safe-region mask (with the probabilities of layers noted); the horizontal color band depicts the replanning time associated with the same ERI (darker = longer). As ERI increases, the distribution becomes more diffuse and the replanning time shortens; as ERI decreases, the distribution concentrates and the replanning time lengthens. This illustrates how a single ERI co-regulates dispersion and timing while the subgoal distribution stays within the safety-consistent corridor.

Benchmark scenes in Arena-3D (left→right, top→bottom): 

• Scene 1, map5 gazebo (sparse, structured layout); 

• Scene 2, factory gazebo (large industrial floor with bulky static occluders); 

• Scene 3, aws house gazebo (multi-room indoor environment with narrow passages); 

• Scene 4, small warehouse gazebo (compact, crowded warehouse, the most challenging one). 

The four scenes form an increasing gradient of geometric and dynamic difficulty used throughout the evaluation.

Quantitative Results

Robustness Curves

Ablation Studies

Ablation Studies on ERI Features