While contemporary deep learning methods are frequently computationally costly, traditional local planners like the Dynamic Window Approach (DWA) are essentially constrained by their purely geometric, “socially blind” nature. This research introduces Semantic-DWA, a unique, lightweight, and interpretable framework that closes this gap by adding a critical layer of semantic knowledge to the traditional DWA. Our methodology utilizes a perception function to categorize obstacles as “person,” “pet,” or “object” and implements a social disqualification rule that treats class-specific proxemic boundaries as hard constraints. Evaluated in a Python-based 2D simulator, comparative results demonstrated that while the standard DWA led to multiple collisions and proxemic violations, the Semantic-DWA completed all runs with zero collisions, maintaining distinct safe clearances such as 1.00 m for persons and 2.08 m for pets. This study indicates that meaningful social intelligence can be added to proven local planners through minimal extensions, offering a verifiable and predictable solution for safer human-robot coexistence.
Ashwath et al. (Wed,) studied this question.