Fast Action Generation via Knowledge Distillation with Flow Matching for Social Navigation

Mobile robot navigation in dynamic environments that contain pedestrians is one of the key challenges in the development of autonomous mobile service robots. This field, known as social navigation, has seen significant research progress using reinforcement learning approaches. In recent years, numerous diffusion-based reinforcement learning methods capable of generating diverse actions have been proposed. However, compared to conventional reinforcement learning approaches, the diffusion model’s slow generation process presents a significant barrier to real-time processing. To address this, we propose a method for knowledge distillation of conditional diffusion models by combining Gaussian Prior with Flow Matching to enable faster action generation in dynamic environments. Experiments using a crowd navigation benchmark in simulation environments demonstrate that a significant reduction of the time required for action generation is possible while maintaining nearly the same performance as teacher models.