Ego2Flow: Benchmarking the Traffic-System Impact of Autonomous Driving on Maps Distilled from Real-World Roads

Yu Dong, Weiying Shen, Hangyu Li, Shu Han, Hong Li, Xiangbo Gao, Keshu Wu, Zhengzhong Tu, Soyoung Ahn, Zhiyuan Liu, Xiaopeng LiarXiv 2026 · arXiv

TL;DR

Ego2Flow is a new closed-loop benchmark designed to evaluate how autonomous driving (AD) policies impact the broader traffic system, moving beyond simple "ego-vehicle" safety metrics. By introducing a novel map-distillation framework that converts real-world Argoverse 2 data into high-fidelity CARLA simulator environments, it reveals a critical insight: a driving policy that is "safe" for an individual car can still be a disaster for overall traffic efficiency and stability.

What Problem Does This Paper Address?

Most autonomous driving research focuses on the ego-vehicle: Does the car reach its destination? Does it avoid collisions? While vital, this "ego-centric" view ignores the collective impact of AD on urban infrastructure.

In the real world, vehicles are part of a fluid system. A single car’s aggressive braking can trigger a "phantom traffic jam" (perturbation propagation) that delays dozens of vehicles behind it. Current benchmarks often fail to capture these interactions because they use open-loop datasets or simplistic simulations. The authors argue that as we move toward widespread deployment, we must evaluate how AD policies affect traffic throughput, string stability (whether speed fluctuations grow or shrink as they move back through a line of cars), and even total energy consumption.

Figure 1. Ego2Flow: From ego-vehicle evaluation to system-level traffic-flow evaluation. Motivation. Strong ego-vehicle performance does not necessarily translate into safer or more stable traffic flow. The schematic illustrates one possible situation: speed disturbances may amplify along a platoon containing autonomous vehicles (AVs), increasing delays for following vehicles, while remaining unamplified in a human-driven vehicle (HV) platoon. Such collective effects are not captured by ego-vehicle performance alone. Benchmark. Ego2Flow addresses this evaluation gap by assessing the traffic-system impact of end-to-end driving policies in closed-loop simulation. It provides six CARLA maps distilled from real-world roads and spatially aligned with the corresponding Argoverse 2 (AV2) regions, supporting evaluation of traffic efficiency, safety, stability and estimated energy consumption across diverse urban road environments.

Methods

The core technical contribution is an agent-guided map distillation framework. Instead of trying to reconstruct every leaf on every tree, the framework focuses on "decision-relevant" features—the geometry and visual cues that actually dictate how a car (or an AI) drives.

Key components include:

  • Traffic Information Graph (TIG): A relational map that links road elements (like a driveway) to source evidence (LiDAR/images) and construction tasks. This ensures that the generated 3D environment isn't just a visual shell but a functionally connected road network.
  • Iterative Feedback Loop: The system uses "Static Reviewers" to check visual alignment with real-world logs and "Dynamic Reviewers" to ensure that AI agents can actually navigate the generated maps without getting stuck or failing due to geometry errors.
  • Closed-Loop Simulation: Unlike "open-loop" testing (where a model predicts what to do on a pre-recorded video), Ego2Flow runs in the CARLA simulator. This allows the actions of the AD policy to change the environment, which in turn changes the observations for the next time step.
Figure 2. Agent-guided map distillation. Aligned AV2 observations and geographic data provide the geometry and evidence for con- struction. The traffic information graph links road elements, evidence, tasks and their simulator realizations. The construction agent uses these relations to constrain road completion and appearance generation. Static and dynamic reviewers return feedback for local refinement. The driveway example connects a curb cut and apron to the evidence and mesh needed to preserve vehicle access.

Datasets

The authors utilized the Argoverse 2 (AV2) dataset as their primary source of real-world road geometry, LiDAR, and imagery.

From this, they distilled and released six new CARLA maps covering diverse urban environments:

  1. Austin
  2. Detroit
  3. Miami
  4. Palo Alto
  5. Pittsburgh
  6. Washington, D.C.

These maps are spatially aligned with the original AV2 regions, allowing researchers to perform "paired" evaluations—comparing how a model performs under real-world recorded observations versus simulated ones.

Key Insights and Ideas

  • The Ego-System Gap: The most striking finding is that strong ego-vehicle performance does not guarantee a healthy traffic system. A model might be excellent at avoiding crashes but terrible at maintaining a steady flow, leading to "stop-and-go" waves.
  • Distillation > Reconstruction: By prioritizing lane topology and visibility over high-fidelity aesthetic details, the authors achieved higher "planning fidelity." In other words, an AI model's planned path in the distilled map was closer to its path in the real world compared to other more complex reconstruction methods.
  • Rule-Based Baselines are Hard to Beat: Surprisingly, classic rule-based controllers (like the Intelligent Driver Model, or IDM) often outperformed state-of-the-art end-to-end neural networks in system-level metrics like throughput and stability. This suggests that current end-to-end models are not yet optimized for "social" driving behaviors.

Key Results

  • Map Fidelity: The Ego2Flow framework achieved a Mean Average Displacement Error (ADE) of 0.7966m for planning outputs, which is 9.3% lower than the next-best construction method (Cam2Sim).
  • Policy Benchmarking: In a comprehensive evaluation of five different driving policies, the IDM-based controller achieved the highest total system score (71.804). Among end-to-end models, LEAD performed best (54.215), but still lagged significantly behind the rule-based baselines in efficiency and safety scores.
  • Stability Metrics: The benchmark successfully quantified "perturbation propagation." It showed that certain policies amplify speed disturbances (making traffic worse), while others can act as "dampers" to smooth out flow.
Table 3. System-level performance of representative driving policies on Ego2Flow. Arrows indicate the preferred direction, and bold marks the best value in each column.

Limitations and Future Work

The authors acknowledge several areas for growth:

  • Visual Complexity: While the distillation captures essential cues, it may still simplify some complex visual "edge cases" found in the real world.
  • Urban Focus: The current maps are primarily urban. Extending this to high-speed freeway environments would introduce different traffic dynamics like lane-changing at scale.
  • Policy Optimization: This paper provides the benchmark, but the next step is using Ego2Flow to train new policies. Future work could involve reinforcement learning where the reward function includes system-level efficiency and stability, not just ego-safety.

Summary updated

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.
BibTeX
@misc{dong2026ego,
  author = {Yu Dong and Weiying Shen and Hangyu Li and Shu Han and Hong Li and Xiangbo Gao and Keshu Wu and Zhengzhong Tu and Soyoung Ahn and Zhiyuan Liu and Xiaopeng Li},
  title = {{Ego2Flow: Benchmarking the Traffic-System Impact of Autonomous Driving on Maps Distilled from Real-World Roads}},
  year = {2026},
  eprint = {2611.17720},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17720},
}
Figure