Ego2Flow: Benchmarking the Traffic-System Impact of Autonomous Driving on Maps Distilled from Real-World Roads
Organizations: University of Wisconsin–Madison · Southeast University · Texas A&M University–College Station
Abstract
End-to-end autonomous driving is predominantly evaluated through ego-vehicle safety and task completion. Its deployment in urban traffic, however, creates sustained interactions among vehicles whose collective effects remain insufficiently evaluated. We introduce Ego2Flow, a closed-loop benchmark for assessing the traffic-system impact of autonomous driving on maps distilled from real-world cities. To construct realistic 3D urban environments for sustained closed-loop traffic simulation, we develop an agent-guided map distillation framework that integrates real-world maps and observations with complementary geographic information. Through an iterative feedback mechanism and a traffic information graph that jointly guide scene generation, the framework preserves lane topology, critical road geometry and decision-relevant visual conditions while simplifying nonessential details and completing unobserved regions. Using our proposed framework, we generate and publicly release six CARLA maps from the Argoverse 2 dataset, covering the corresponding real-world regions and enabling paired comparisons of driving behavior under real and simulated observations. On these maps, Ego2Flow evaluates how autonomous driving policies affect traffic efficiency, safety, stability and energy consumption across diverse road structures. Compared with existing construction methods, our framework achieves the lowest average displacement error between planning outputs from paired real and simulated observations. We further benchmark representative driving policies to characterize their system-level impact on urban traffic. Project Page: https://github.com/Fisher-007/Ego2Flow.
Figures & tables
| Method | City-wise displacement error | Mean ∗ | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Palo Alto | Detroit | Miami | Austin | Pittsburgh | Washington, D.C. | |||||||||
| ADE | FDE | ADE | FDE | ADE | FDE | ADE | FDE | ADE | FDE | ADE | FDE | ADE | FDE | |
| CARLA Digital Twin Tool [ 4 ] | 0.7432 | 2.0995 | 1.1145 | 2.9851 | 1.2408 | 2.9613 | 1.2948 | 3.5720 | 1.1814 | 3.2558 | 0.9320 | 2.3710 | 1.0844 | 2.8741 |
| UrbanWorld2.0 [ 31 ] | 0.6937 | 2.0092 | 1.0826 | 2.9120 | 1.2109 | 2.9682 | 1.2728 | 3.5452 | 1.1240 | 3.1217 | 0.7031 | 1.8364 | 1.0145 | 2.7321 |
| MDrive [ 8 ] | 0.7230 | 2.1126 | 0.7531 | 2.0016 | 1.0828 | 2.6757 | 1.0315 | 2.8889 | 1.2353 | 3.5230 | 0.6074 | 1.5759 | 0.9055 | 2.4630 |
| TRACE † [ 28 ] | 0.7842 | 2.2787 | 0.8755 | 2.3178 | 1.0969 | 2.6450 | 1.6192 | 4.5568 | 1.1992 | 3.4441 | 0.7147 | 1.8290 | 1.0483 | 2.8452 |
| Variant | Framework components | Results (5 s) | |||
|---|---|---|---|---|---|
| Graph | Static feedback | Dynamic feedback | ADE | FDE | |
| w/o graph | — | 0.3542 | 0.9363 | ||
| w/o static feedback | — | 0.3706 | 0.9918 | ||
| w/o dynamic feedback | — | 0.3634 | 0.9791 | ||
| Full Ego2Flow | 0.3465 | 0.9285 | |||
| Method | Efficiency | Safety | Stability | Energy | Overall | |||||
|---|---|---|---|---|---|---|---|---|---|---|
| delay score | throughput score | collision score | low-TTC score | braking score | speed SD score | peak loss score | energy score | failure fraction | total score | |
| Normalized scores: | ||||||||||
| IDM-based [ 30 ] | 0.7510 | 0.8868 | 0.8733 | 0.8431 | 0.8150 | 0.5479 | 0.5242 | 0.5031 | 0.0000 | 71.804 |
| Traffic Manager [ 5 ] | 0.7048 | 0.8082 | 0.9192 | 0.5944 | 0.6038 | 0.5231 | 0.4862 | 0.5295 | 0.0000 | 64.615 |
| LEAD [ 23 ] | 0.5671 | 0.5608 | 0.8360 | 0.7649 | 0.4480 | 0.5083 | 0.1528 | 0.4994 | 0.0104 | 54.215 |
| UniAD [ 14 ] | 0.4850 | 0.3349 | 0.7687 | 0.7045 | 0.3347 | 0.6107 | 0.2355 | 0.4648 | 0.0000 | 49.235 |
Autonomous Driving Benchmarks
Autonomous driving benchmarks evaluate how well vehicles perceive, predict, plan, and act safely across varied roads and conditions.
See all 158 papers