Flow Head Rollout: Scalable CAV Coordination

• Flow Head Rollout is a framework that integrates macroscopic traffic modeling with adaptive, truncated MPC to coordinate connected and autonomous vehicles in congested highway settings.
• It employs sequential agent-by-agent optimization and horizon truncation, reducing computational complexity while ensuring throughput and closed-loop stability.
• Empirical results show a 17% reduction in travel time versus no control and a 22.9% decrease in CPU cost compared to centralized MPC.

Flow Head Rollout denotes a class of multiagent sequential optimization algorithms developed for large-scale coordination of connected and autonomous vehicles (CAVs) in congested highway environments under mixed autonomy. It unifies macroscopic traffic flow modeling, distributed model predictive control (DMPC), and agent-specific adaptive planning through truncation of rollout horizons. The approach exploits explicit representations of CAVs as moving bottlenecks over discretized highway segments and supports rigorous guarantees on both performance and closed-loop stability. Flow Head Rollout distinguishes itself by deploying horizon truncation and agent-by-agent solution scheduling to avoid the exponential complexity of full-horizon, full-joint optimization, thereby achieving state-of-the-art throughput and scalability in real-time mixed traffic scenarios (Liu et al., 26 Aug 2025).

1. Macroscopic Density Evolution and Mixed Autonomy Modeling

Flow Head Rollout algorithms ground their system model in continuum traffic theory, specifically the Lighthill–Whitham–Richards (LWR) partial differential equation for vehicle density $\rho(x,t)$: $$\partial_t \rho + \partial_x f(\rho) = 0,\qquad f(\rho) = \rho\,v(\rho), \quad v(\rho) = V\left(1 - \frac{\rho}{R}\right)$$ where $V$ is free-flow speed and $R$ is the jam density. Each CAV $i$ follows the ODE: $\dot y_i(t) = \min\{u_i(t),\, v(\rho(y_i^+, t))\}$ with $u_i(t)$ its control input (velocity setpoint), subject to local downstream flow limits.

For algorithmic deployment, spatial discretization via the cell-transmission model (CTM) partitions the road into cells, yielding a vectorized update

$$\bm\rho^{k+1} = \bm\rho^k + \mathbf{A}^k\,\mathbf{O}^k + \mathbf{B}^k + \mathbf{C}_{\mathbf{u}}^k$$

where $\bm\rho^k$ comprise cell densities, $\mathbf{O}^k$ marks cells with CAV bottlenecks, $\mathbf{A}^k$ and $\mathbf{B}^k$ are discretized flux operators, and $\mathbf{C}_{\mathbf{u}}^k$ encodes agent control effects. The model rigorously reconstructs upstream and downstream densities at moving bottlenecks by solving

$$f(\rho) - u_i\rho = \frac{\alpha R}{4V}(V-u_i)^2\qquad \alpha = \frac{W-1}{W}$$

with $W$ the number of lanes (Liu et al., 26 Aug 2025).

2. Multiagent Truncated Rollout Strategy

Core to Flow Head Rollout is the adaptive truncation of each agent’s optimization horizon within a finite-horizon Model Predictive Control (MPC) framework. Formally, each agent $i$ solves: $$\min_{\bm\mu_i^k} J_i(\bm\xi_i^k, \bm\mu_i^k) = \sum_{t=0}^{N-1}\sum_{j=0}^J \rho_{i,j}^{k+t}\, \Delta t\, \Delta x$$ subject to the discretized density evolution, terminal set and contraction constraint: $$V_i(\bm\xi_i^k, \bm\mu_i^k) \leq \eta_i(\bm\xi_i^k, \lambda_i)$$ where $V_i$ is an ISS-Lyapunov cost and $\eta_i$ is a contraction function enforcing stability. Rather than solving the full $N$-step MPC problem for every CAV, the previous solution vector $\bm\mu_i^{k*,p-1} = [\mu_{f},\mu_{r}]$ is split, and only the first $M_i^J$ decision variables are optimized: $$M_i^J = M_\mathrm{min} + (N - M_\mathrm{min})\, \frac{\max\{J_i^{k,LB}, \min\{J_i(\cdot), J_i^{k,UB}\}\} - J_i^{k,LB}}{J_i^{k,UB} - J_i^{k,LB}}$$ ensuring computationally efficient adaptation of rollout depth according to real-time cost bounds (Liu et al., 26 Aug 2025).

3. Sequential Agent-by-Agent Optimization

Instead of parallel coordination, Flow Head Rollout imposes a strict agent-by-agent sequential update schedule. For each round, CAVs are ordered (e.g., by an "impact ranking") and each agent in turn solves its personal truncated-horizon MPC problem while using the most recently updated strategies of predecessors and a fixed or placeholder model for successors. After one full permutation, a convergence check on the terminal agent's cost determines if a new round is necessary. This scheme ensures computational tractability, implicit negotiation, and monotonic non-increase of the cost functions: $$J_i(\bm\xi_i^{k*,p-1}, \bm\mu_i^{k*,p-1}) \ge J_i(\bm\xi_i^{k*,p}, \bm\mu_i^{k*,p})$$ guaranteeing local—and often global—optimality in coordination (Liu et al., 26 Aug 2025).

4. Stability, Input-to-State Guarantees, and Theoretical Properties

The approach is equipped with formal stability and recursive feasibility properties. Each agent admits an ISS Lyapunov function $V_i$ such that for appropriate terminal sets and local feedback, the closed-loop policy ensures: $$V_i(k+1) - V_i(k) \leq - (1-\lambda_i)q \|\rho_i^k\|^2 < 0$$ for contraction parameter $\lambda_i \in [0,1)$, ensuring robust convergence and constraint satisfaction throughout real-time control. The sequential truncation update scheme strictly improves or leaves unchanged all agents' values relative to parallel DMPC or centralized MPC, supporting provable safety and performance (Liu et al., 26 Aug 2025).

5. Empirical Performance and Computational Scalability

Deployment on the Shanghai Hujin-Expressway bottleneck with $15\%$ CAV market penetration and $2,000$ veh/h inflow yields:

• No control: $T_{\rm avg}=191.54\, {\rm s}$
• Centralized MPC: $165.94\, {\rm s}$ ($-13.37\%$)
• DMPC: $166.20\, {\rm s}$ ($-13.23\%$)
• Flow Head Rollout: $158.43\, {\rm s}$ ($-17.29\%$ vs no control; $-4.55\%$ vs MPC)

Computational complexity per agent drops from $O(S^{(I+1)N})$ (centralized) and $O((I+1)S^N)$ (DMPC) to

$O\left(\sum_{i=0}^I S^{M_i^J}\right)$

where $M_i^J \ll N$ is typically achieved in practice. This yields a $22.9\%$ reduction in CPU time vs centralized MPC, while matching the scalability of DMPC and exceeding its throughput, with only $5.7\%$ more computational cost (Liu et al., 26 Aug 2025).

6. Algorithm Summary and Implementation Notes

The Flow Head Rollout procedure encapsulates:

• Real-time density state observation and cell assignment,
• Optional agent ordering by marginal cost impact,
• Initialization of current and previous control sequences,
• Inner iterative cycle: each agent solves a truncated MPC on revised densities, using the latest policies of other agents,
• Convergence monitored via cost improvement,
• Application of the first control input and advancement of system state.

The method accommodates entry and exit of CAVs, changes in traffic conditions, and is directly applicable for large-scale deployments with minimal architecture modification.

7. Relationship to Broader Multiagent Rollout Techniques

Flow Head Rollout extends the multiagent rollout paradigm for highway decongestion and throughput improvement—see also agent-by-agent policy iteration frameworks in Dec-POMDP traffic flow (Liu et al., 2024). In contrast to earlier MARollout/MARollout approaches, which use fixed or agent-wise horizons and may suffer exponential scaling, the adaptive truncation and explicit moving bottleneck integration in Flow Head Rollout enable both rigorous theoretical guarantees and best-of-class empirical results in realistic bottleneck scenarios (Liu et al., 26 Aug 2025, Liu et al., 2024).