Semantic Obstacle Avoidance Method for Humanoid Service Robots Based on Visual Language Model and Proximal Policy Optimization

Traditional geometry-based obstacle avoidance methods for humanoid service robots lack semantic understanding, leading to rigid behaviors in dynamic unstructured human-coexistence environments. This paper proposes a semantic obstacle avoidance method integrating visual language models and proximal policy optimization (PPO). First, Grounded SAM2 realizes open-vocabulary recognition and segmentation of environmental obstacles while acquiring their semantic categories. Second, PPO with reward shaping trains obstacle avoidance strategy primitives to construct a behavior library covering dense environment navigation and risk-aware planning. Third, a semantic policy scheduler dynamically matches optimal strategies based on real-time semantic information. Finally, Lagrange interpolation smooths control commands for natural and safe navigation. Experiments show the method significantly improves obstacle avoidance success rate and human-like behavior performance.