ConsistencyPlanner: Real-time Planning with Fast-Sampling Consistency Models

The article introduces Consistency Planner, a novel real-time planning framework for autonomous driving that leverages fast-sampling consistency models to generate diverse future trajectories efficiently. Key innovations include efficient multimodal sampling to enhance real-time exploration of driving actions and an attention-enhanced decoder for integrating heterogeneous input features, resulting in improved safety metrics in dynamic environments. This advancement addresses the limitations of previous learning-based approaches, making it significant for practitioners focused on developing robust autonomous driving systems.