Aligning Microscopic Vehicle and Macroscopic Traffic Statistics: Reconstructing Driving Behavior from Partial Data

TL;DR

This work addresses learning driving policies from partially observed data by bridging microscopic demonstrations and macroscopic traffic statistics. It introduces a two-stage macro–micro framework: a generator $G_\phi$ completes hidden microscopic states from partial observations and macro descriptors, and a shared policy $\pi_\theta$ is trained to maximize a trajectory-level objective that combines microscopic imitation and macroscopic alignment, using $J(\theta) = \mathbb{E}_{\tau} [ r_{\mathrm{micro}}(\tau;\theta) + \eta\, r_{\mathrm{macro}}(\tau;\theta) ]$. In a ring-road case study, the learned policy preserves target mean speed and spacing while producing realistic microscopic car-following trajectories, demonstrating macro–micro consistency and safe coordination with human drivers at scale. This approach reduces reliance on densely labeled micro data and provides a scalable path toward safe, human-aligned autonomous driving in partially observed real-world environments.

Abstract

A driving algorithm that aligns with good human driving practices, or at the very least collaborates effectively with human drivers, is crucial for developing safe and efficient autonomous vehicles. In practice, two main approaches are commonly adopted: (i) supervised or imitation learning, which requires comprehensive naturalistic driving data capturing all states that influence a vehicle's decisions and corresponding actions, and (ii) reinforcement learning (RL), where the simulated driving environment either matches or is intentionally more challenging than real-world conditions. Both methods depend on high-quality observations of real-world driving behavior, which are often difficult and costly to obtain. State-of-the-art sensors on individual vehicles can gather microscopic data, but they lack context about the surrounding conditions. Conversely, roadside sensors can capture traffic flow and other macroscopic characteristics, but they cannot associate this information with individual vehicles on a microscopic level. Motivated by this complementarity, we propose a framework that reconstructs unobserved microscopic states from macroscopic observations, using microscopic data to anchor observed vehicle behaviors, and learns a shared policy whose behavior is microscopically consistent with the partially observed trajectories and actions and macroscopically aligned with target traffic statistics when deployed population-wide. Such constrained and regularized policies promote realistic flow patterns and safe coordination with human drivers at scale.

Figures (3)

• Figure 1: Partially observed traffic setting. A small subset of instrumented vehicles (blue) provides microscopic trajectories and actions, while the remaining vehicles (red) are unobserved. Roadside sensors (e.g., cameras/detectors) supply macroscopic statistics for the whole stream
• Figure 2: Two-stage framework. Top (Step 1): the generator $G_\phi$ takes partial snapshots $s_0^{\mathrm{obs}}$, completes hidden vehicles to form $\hat{S}_0$, and is trained with the generator loss $\mathcal{L}_{\mathrm{gen}}(\phi)$. Bottom (Step 2): episodes start from $\hat{S}_0$; the shared policy $\pi_\theta$ rolls out the environment and is updated using a trajectory-level score combining microscopic imitation $r_{\mathrm{micro}}$ and macroscopic alignment $r_{\mathrm{macro}}$. In each simulation, observed vehicles follow their ground-truth trajectories, while unobserved vehicles are controlled by the learned policy.
• Figure 3: Leader–follower velocity profiles in two scenarios. Top: leader accelerates 11m/s→13 m/s; bottom: leader decelerates 13m/s→11 m/s. In these plots, the velocity trajectories are drawn in green (leader), blue (ground-truth IDM follower), and orange (our learned policy follower), respectively. Across runs, the policy’s average profile follows the ground-truth trend.