Socially Adaptive Navigation via Recursive Generative Knowledge Learning and Transfer

The increasing deployment of mobile service robots demands enhanced capabilities in learning and transfer of navigation behavior for complex human-robot interactions. Current social navigation approaches often treat pedestrians merely as dynamic obstacles, neglecting historical scene information and long-term social behavior prediction, resulting in memory-deficient navigation. We propose a recursive generative framework based on inverse deep reinforcement learning. By integrating recurrent neural networks and innovatively proposing the pedestrian convergence probability value as a novel social behavior feature, our RNN-RRT planner integrates recurrent neural networks to simultaneously optimize pedestrian social costs, path length, and obstacle avoidance while maintaining memory through historical feature integration. Experimental results demonstrate that the trained RNN-RRT planner achieves effective transfer learning in novel environments, accurately predicting pedestrian convergence, and significantly improving both path generalization and human-like navigation characteristics.