FlowAD: Ego-Scene Interactive Modeling for Autonomous Driving

Mingzhe Guo^1,2, Yixiang Yang¹, Chuanrong Han¹, Rufeng Zhang², Shirui Li², Ji Wan², Zhipeng Zhang^{1 ✉}

¹ AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University, ² Baidu Inc.

^✉ corresponding author: zhipeng.zhang.cv@outlook.com

Accepted to ICLR 2026!

Table of Contents

• Introduction
• Performance Highlights
  • nuScenes Open-Loop Evaluation
  • Qualitative Results - Perception
  • Bench2Drive Closed-Loop Evaluation
• Getting Started
• Citation
• Acknowledgements
• License

Introduction

This repository contains the official implementation of FlowAD, a novel ego-scene interactive modeling framework for autonomous driving. Unlike traditional approaches that treat each timestamp in isolation, FlowAD explicitly models the feedback of ego-vehicle motion to future observations, fundamentally improving the understanding of the driving process and enhancing planning capabilities.

The architecture of our FlowAD structured around three core components: 1) Ego-guided scene partition. 2) Spatial and temporal flow prediction. 3) Task-aware enhancement.

Inspired by human perception, FlowAD represents ego-scene interaction as scene flow relative to the ego-vehicle, capturing relative motion as learnable scene flow within the latent feature space. This enables modeling ego-motion feedback using existing log-replay datasets without requiring complex scenario simulations.

Key Achievements:

• 19% collision rate reduction over SparseDrive on nuScenes
• 60% FCP (our proposed metric) improvement (1.39 frames) on nuScenes validation set
• 51.77 driving score on Bench2Drive closed-loop evaluation
• Demonstrated generality across perception, end-to-end planning, and VLM analysis

Performance Highlights

nuScenes Open-Loop Evaluation

Our method achieves significant improvements across multiple tasks:

Method	Backbone	Detection		Tracking		Motion Prediction		Planning		FCP↓
		mAP↑	NDS↑	AMOTA↑	AMOTP↓	minADE↓	minFDE↓	Avg.L2 (m)↓	Avg.Col↓
UniAD	ResNet101	0.380	0.498	0.359	1.320	0.71	1.02	0.69	0.12	2.96
SparseDrive	ResNet101	0.496	0.588	0.501	1.085	0.60	0.96	0.58	0.06	2.30
FlowAD (Ours)	ResNet101	0.523	0.605	0.518	1.040	0.56	0.93	0.52	0.05	0.91

FCP (Frames before Correct Planning): Lower is better. FlowAD achieves 1.39 frames improvement (48% reduction) over SparseDrive and 2.03 frames improvement (60% reduction) over baseline methods.

Qualitative Results - Perception

FlowAD demonstrates superior detection of occluded objects, small targets, and dense scenes through learned scene flow dynamics.

Bench2Drive Closed-Loop Evaluation

FlowAD achieves 51.77 driving score, demonstrating robust closed-loop performance.

Getting Started

Quick Start

1. Clone this repository:

git clone https://github.com/your-repo/FlowAD.git
cd FlowAD

2. Navigate to the desired sub-project:

cd SparseDrive-Flow 

3. Follow the sub-project's README for environment setup and training/evaluation.

Citation

If you find this work useful in your research, please consider citing:

@inproceedings{FlowAD2026,
  title={FlowAD: Ego-Scene Interactive Modeling for Autonomous Driving},
  author={Anonymous},
  booktitle={Under review as a conference paper at ICLR 2026},
  year={2026},
  url={https://openreview.net/pdf?id=m4JpoJRgAr}
}

Acknowledgements

This work builds upon several excellent open-source projects:

• SparseDrive - End-to-end autonomous driving

License

This project is released under the Apache 2.0 license. Please see the LICENSE files in each sub-project for more details.