LAP: Fast LAtent Diffusion Planner for Autonomous Driving

TL;DR

LAP is a novel latent space planning framework for autonomous driving that significantly reduces inference time and improves planning quality by disentangling high-level intents from low-level kinematics using a VAE-based approach.

Contribution

The paper introduces LAP, a latent space planner that enables fast, high-quality multi-modal driving plans with a single denoising step, addressing latency and low-level detail issues in diffusion models.

Findings

LAP achieves state-of-the-art performance on nuPlan benchmark.

LAP reduces inference time by up to 10x compared to previous methods.

LAP produces high-quality plans in a single denoising step.

Abstract

Diffusion models have demonstrated strong capabilities for modeling human-like driving behaviors in autonomous driving, but their iterative sampling process induces substantial latency, and operating directly on raw trajectory points forces the model to spend capacity on low-level kinematics, rather than high-level multi-modal semantics. To address these limitations, we propose LAtent Planner (LAP), a framework that plans in a VAE-learned latent space that disentangles high-level intents from low-level kinematics, enabling our planner to capture rich, multi-modal driving strategies. To bridge the representational gap between the high-level semantic planning space and the vectorized scene context, we introduce an intermediate feature alignment mechanism that facilitates robust information fusion. Notably, LAP can produce high-quality plans in one single denoising step, substantially reducing computational overhead. Through extensive evaluations on the large-scale nuPlan benchmark, LAP achieves state-of-the-art closed-loop performance among learning-based planning methods, while demonstrating an inference speed-up of at most 10x over previous SOTA approaches.