Context-appropriate Social Navigation in Various Density Construction Environment using Reinforcement Learning

Abstract

Construction environments are often densely populated with multiple resources (e.g., workers, equipment, and materials). As an increasing number of mobile robots are expected to coexist and interact with humans at close proximity, it is necessary that these robots are capable of not only avoiding collisions with people but also not disturbing human work and deteriorating human comfort. Failing to maintain a proper social space can lead to fatal accidents and inefficiency. To accommodate this need, this study aims to develop a social navigation model that enables robots to navigate in a contextually compliant manner. We created a simulation environment where robot agents can learn socially and contextually aware policies using reinforcement learning. The results showed that the agent was able to secure the respective minimum separation distance for different types of workers while achieving similar overall performance in contrast to baseline models which often violated the work-related proxemic considerations. This finding will contribute to building future construction mobile robots with social intelligence which are capable of understanding the context of the workplace and adapting to appropriate behaviors accordingly.