README.md

April 26, 2026 · View on GitHub

Mimir: Hierarchical Goal-Driven Diffusion with Uncertainty Propagation for End-to-End Autonomous Driving
Zebin Xing^1*, Yupeng Zheng¹, Qichao Zhang¹, Zhixing Ding^1,2, Pengxuan Yang¹, Songen Gu³, Zhongpu Xia¹, Dongbin Zhao¹
¹ Institue of Automation, Chinese Academy of Sciences, ² China University of Geosciences, ³ Fudan University

IEEE Robotics and Automation Letters (RAL), 2025

Mimir is a hierarchical goal-driven diffusion model for end-to-end autonomous driving. It improves upon GoalFlow by explicitly modeling goal uncertainty and accelerating goal inference, enabling more robust and efficient planning. With a ResNet-34 backbone, Mimir achieves 89.3 PDMS and 34.6 EPDMS, demonstrating strong performance in both accuracy and efficiency.

News

26 Apr, 2026: We released the code on navtest.
8 Dec, 2025: We released our paper on arXiv. Code is coming soon.
17 Nov, 2025: Mimir was accepted at RAL !

To Do

[x] Code for training
[x] Code for validation, Weight of model
[x] Initial repo & main paper

Introduction

End-to-end autonomous driving has shown strong potential with the aid of high-level guidance, yet its performance is often limited by inaccurate guidance and high computational cost. To address these issues, we propose Mimir, a hierarchical dual-system framework that generates robust trajectories with uncertainty-aware goal modeling. Specifically, Mimir models goal uncertainty using a Laplace distribution to improve robustness and introduces a multi-rate guidance mechanism to accelerate high-level inference by predicting extended goal points in advance. Experiments on the challenging Navhard and Navtest benchmarks show that Mimir outperforms prior methods by 20% in EPDMS, while achieving 1.6× faster inference for the high-level module without sacrificing accuracy. Code and models will be released to facilitate future research.

Getting Started

We provide essential model weights and intermediate files (e.g., navigation points and uncertainty information). Mimir employs two separate models to generate uncertain navigation information and planning trajectories respectively.

📦 Model Zoo

Agent	Size	Link	Description
mimir_unc	724MB	Download	Generates navigation points with uncertainty modeling
mimir	746MB	Download	Plans trajectories based on uncertain navigation information

📋 Data Preparation

Before proceeding, please complete the following prerequisites:

Download NAVSIM Dataset - Follow the NAVSIM installation guide to download the OpenScene dataset and nuPlan maps
Generate Metric Cache - Create evaluation cache containing necessary metrics
Generate Feature Cache - Prepare training features for model training

📊 Evaluation

The evaluation process requires the Metric Cache generated in the Data Preparation step, which stores essential evaluation information.

Our evaluation pipeline consists of two stages:

Step 1: Generate navigation points and uncertainty information

bash scripts/run_generate_mimir_unc.sh

Step 2: Generate trajectories and evaluate performance

bash scripts/evaluation/run_mimir.sh

🎓 Training

We employ a two-stage training strategy to train mimir_unc and mimir agents separately:

Train mimir_unc (uses GoalFlow predicted goal points as initialization)
```
bash scripts/training/run_mimir_unc_training.sh
```
Generate navigation and uncertainty data for training set
```
bash scripts/run_generate_mimir.sh
```
Train mimir agent with generated navigation information
```
bash scripts/training/run_mimir_training.sh
```

⚡ Asynchronous Inference

💡 Note: Intermediate navigation and uncertainty data are already provided. You can skip any stage and start development directly.

Cross-Benchmark Compatibility:

Both navhard and navtest benchmarks use identical models for testing
For navhard testing, you can directly use NAVSIM v2.x development library
Simply migrate the contents from navsim/agents/mimir/ and navsim/planning/script/config/common/agent/mimir_agent*.yaml for development

Visualization

Comparison with Other Methods

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.

🌊 GoalFlow - Goal point generation and flow-based prediction
🚗 DiffusionDrive - Diffusion-based trajectory planning framework
🗺️ NAVSIM - Simulation platform and benchmark

We sincerely thank all contributors for their valuable work that helped shape this research.

Citation

If you find Mimir useful, please consider giving us a star 🌟 and citing our paper with the following BibTeX entry.

@ARTICLE{11282450,
  author={Xing, Zebin and Zheng, Yupeng and Zhang, Qichao and Ding, Zhixing and Yang, Pengxuan and Gu, Songen and Xia, Zhongpu and Zhao, Dongbin},
  journal={IEEE Robotics and Automation Letters}, 
  title={Mimir: Hierarchical Goal-Driven Diffusion With Uncertainty Propagation for End-to-End Autonomous Driving}, 
  year={2026},
  volume={11},
  number={2},
  pages={2178-2185},
  keywords={Uncertainty;Trajectory;Predictive models;Autonomous vehicles;Laser radar;Vocabulary;Planning;Feature extraction;Estimation;Artificial intelligence;Learning from demonstration;imitation learning;autonomous vehicle navigation},
  doi={10.1109/LRA.2025.3641129}}

(back to top)