Polite Detouring Strategy for Social Compliance Obstacle Avoidance of Humanoid Service Robots

Humanoid service robots struggle to achieve socially compliant navigation around pedestrians in human-centric environments. Conventional geometric-based obstacle avoidance often leads to rigid, intrusive motions that breach social norms. This paper introduces a novel semantic avoidance method for respectful human-robot interaction. It utilizes vision-language models for pedestrian detection and a reinforcement learning framework based on Trust Region Policy Optimization (TRPO). The core innovation is a carefully designed reward function emphasizing social compliance and motion smoothness, enabling robots to learn polite detouring. Tests on the Pepper robot platform show a 95% success rate in pedestrian interactions, an average minimum social distance of 0.8 m, and a 62.5% reduction in jitter compared to the visibly shaky DWA method. User feedback reflects enhanced comfort, averaging 4.5/5.0. This approach enables humanoid service robots to achieve natural, comfortable, and socially compliant obstacle avoidance in dynamic human settings.