Core Insights - The article presents LAP (LAtent Planner), a framework designed to enhance autonomous driving by decoupling high-level intentions from low-level kinematics, allowing for efficient planning in a semantic space [2][39]. - LAP significantly improves modeling capabilities for complex, multimodal driving strategies and achieves a tenfold increase in inference speed compared to current state-of-the-art methods [1][22]. Background Review - The development of autonomous driving systems has faced challenges in robust motion planning within complex interactive environments, leading to the introduction of LAP to address these issues [2]. Methodology - LAP framework decomposes trajectory generation into two stages: planning in a high-level semantic latent space and reconstructing the corresponding trajectory with high fidelity [8][39]. - The framework utilizes a Variational Autoencoder (VAE) to compress raw trajectory data into a semantic latent space, enhancing the model's focus on high-level driving strategies [10][39]. Experimental Results - LAP achieved superior performance on the nuPlan benchmark, surpassing previous state-of-the-art methods by approximately 3.1 points on the challenging Test14-hard dataset [22][39]. - The inference speed of LAP is significantly improved, requiring only 2 sampling steps to generate high-quality trajectories, compared to 10 steps for previous methods [22][27]. Key Contributions - The framework effectively decouples high-level semantics from low-level kinematics using a VAE, facilitating better interaction between planning and contextual scene information [40]. - The introduction of fine-grained feature distillation bridges the gap between the latent planning space and the vectorized scene context, enhancing model performance [40]. - LAP achieves state-of-the-art closed-loop performance on the nuPlan benchmark while improving inference speed by a factor of 10 [40].