JPPD: Joint Prediction_Planning Diffusion with Differentiable Safety Guidance for Dynamic Obstacle Avoidance in Intelligent Transportation Systems

The paper introduces JPPD, a Joint Prediction-Planning Diffusion framework designed for low-speed autonomous navigation in shared spaces, which integrates trajectory prediction and robot planning into a single conditional trajectory generation problem using a causal Transformer with cross-trajectory attention. Key innovations include differentiable safety potential guidance to replace traditional heuristic methods and conditional flow matching to enhance multimodal trajectory diversity while reducing inference steps. This approach has been validated through various simulations, demonstrating improvements in safety and runtime efficiency compared to conventional prediction-then-planning methods, making it significant for practitioners in intelligent transportation systems.