Mimir: Hierarchical Goal-Driven Diffusion with Uncertainty Propagation for End-to-End Autonomous Driving

TL;DR

<3-5 sentence high-level summary> Mimir introduces a hierarchical dual-system for end-to-end autonomous driving that explicitly models goal-point uncertainty with a Laplace distribution and accelerates high-level guidance via multi-rate extrapolation. The slow guidance module supplies uncertain goal points, which are injected into a fast, diffusion-based planner through a Guidance Injection mechanism, enabling robust trajectory generation. Empirical results on Navhard and Navtest benchmarks show state-of-the-art performance, with about 20% gains in driving score and a 1.6× speedup in high-level inference, demonstrating practical real-time viability. The work highlights the value of separating uncertainty-aware high-level guidance from fast low-level planning to improve safety, efficiency, and robustness in autonomous driving systems.

Abstract

End-to-end autonomous driving has emerged as a pivotal direction in the field of autonomous systems. Recent works have demonstrated impressive performance by incorporating high-level guidance signals to steer low-level trajectory planners. However, their potential is often constrained by inaccurate high-level guidance and the computational overhead of complex guidance modules. To address these limitations, we propose Mimir, a novel hierarchical dual-system framework capable of generating robust trajectories relying on goal points with uncertainty estimation: (1) Unlike previous approaches that deterministically model, we estimate goal point uncertainty with a Laplace distribution to enhance robustness; (2) To overcome the slow inference speed of the guidance system, we introduce a multi-rate guidance mechanism that predicts extended goal points in advance. Validated on challenging Navhard and Navtest benchmarks, Mimir surpasses previous state-of-the-art methods with a 20% improvement in the driving score EPDMS, while achieving 1.6 times improvement in high-level module inference speed without compromising accuracy. The code and models will be released soon to promote reproducibility and further development. The code is available at https://github.com/ZebinX/Mimir-Uncertainty-Driving

Figures (5)

• Figure 1: Compared to existing methods, Mimir assesses the uncertainty of the predicted goal points from the high-level model, enabling safer and more reliable trajectory planning.
• Figure 2: Overview of the Mimir architecture. The high-level model (orange) selects the optimal goal point from a predefined vocabulary $\mathbb{V}$ and estimates its uncertainty using a Laplace distribution to generate guidance $G=\{\bm{\mu},\bm{b}\}$. The multi-rate Guidance further predicts the extended goal point, allowing the system to operate at different rates. The low-level planner (green) fuses the guidance information with perception features through a Guidance Injection module, then generates trajectories via the diffusion planner. Mimir decouples guidance generation from trajectory planning for optimized efficiency.
• Figure 3: Extended Goal Point Extrapolation. (1) Linear Kinematic Extrapolation directly performs linear extrapolation based on historical trajectory points; (2) DAC-Score Guided Kinematic Extrapolation refines the extended goal point prediction by incorporating DAC scores during extrapolation. In the DAC map visualization, red points indicate higher DAC scores approaching 1, while blue points represent lower scores near 0.
• Figure 4: Guidance Injection module.
• Figure 5: Visualization. From top to bottom, we illustrate the capabilities of the four algorithms in Exit-Ramp Scenario, forked-intersection scenario, and roundabout scenario, respectively. The red cross represents the predicted goal point and the size of the yellow area around the goal point represents the level of uncertainty. Mimir leverages uncertainty estimation to mitigate the effects of inaccurate high-level guidance, enabling the generation of safer trajectories.