PreSimNet: A Synergistic Physics-Encoded Deep Learning Framework for Integrated Prediction and Simulation of Car-Following in Mixed-Autonomy Traffic

This is the official repository for our submitted paper: "A Synergistic Physics-Encoded Deep Learning Framework for Integrated Prediction and Simulation of Car-Following in Mixed-Autonomy Traffic". All code and weights files will be made public within one week of the acceptance of the manuscript.

---

Overview

Modeling heterogeneous car-following dynamics is a critical challenge in mixed-autonomy traffic, which comprises both autonomous (AVs) and human-driven vehicles (HVs). Existing research often treats trajectory prediction and behavior simulation as isolated tasks. This creates a trade-off between the short-term numerical accuracy of prediction models and the long-term behavioral consistency of simulation models.

To resolve this conflict, we propose PreSimNet, a novel, physics-encoded deep learning framework that synergistically integrates both tasks within a unified architecture. The core of PreSimNet is a shared encoder that learns type-guided features for all four interaction types (AV-AV, AV-HV, HV-AV, HV-HV). This common representation then informs two specialized decoders: an open-loop head for high-accuracy trajectory prediction, and a closed-loop Mixture-of-Experts (MoE) module that generates dynamic parameters for physics-based models to enable high-fidelity simulation.

Our experiments show that PreSimNet significantly outperforms specialized baselines in both prediction and simulation tasks. Long-term comparative experiments further validate the synergistic design, showing that the closed-loop simulation maintains behavioral realism where open-loop prediction fails due to catastrophic error accumulation.

Framework

The core idea of PreSimNet is to bridge the gap between trajectory prediction and behavior simulation, allowing them to benefit from each other.

Figure: (a) Traditional trajectory prediction models focus on short-term positional accuracy. (b) Traditional behavior simulation models prioritize long-term dynamic consistency. (c) The proposed PreSimNet synergistically integrates both tasks through a shared feature encoder and specialized decoders, achieving both high-accuracy prediction and high-fidelity simulation.

The framework consists of three main components:

1. Trajectory Feature Learning: Acts as a shared encoder using Mamba and Transformer blocks to extract deep spatio-temporal features from historical trajectories and explicitly predict the car-following interaction type.
2. Open-Loop Prediction: A specialized decoder for direct, multi-step trajectory forecasting. It takes the shared features to generate future positions with the highest possible short-term accuracy.
3. Closed-Loop Simulation: Another specialized decoder for behavior simulation. It utilizes a physics-encoded Mixture-of-Experts (MoE) to dynamically generate parameters for classic car-following models (IDM and ACC) based on the shared features. It then iteratively simulates the vehicle's future velocity and gap in a closed-loop manner, ensuring long-term behavioral realism and dynamic feasibility.

Key Results

PreSimNet demonstrates state-of-the-art performance across multiple benchmarks.

1. Trajectory Prediction & Type Classification

PreSimNet achieves an average RMSE of 0.258 m in short-term trajectory prediction and a classification accuracy of 99.23% for car-following types, outperforming all baselines.

Model	Position RMSE (m)	Classification Accuracy (%)	#Params	Inf. Time (ms/batch)
	Avg.	Avg.
Seq2Seq	0.304	97.12%	185,413	2.4
Transformer	0.295	98.62%	246,085	2.9
CS-LSTM	0.380	95.68%	247,525	4.9
STDAN	0.278	97.10%	336,165	6.3
BAT	0.272	98.39%	417,605	5.1
HLTP	0.280	98.96%	527,882	7.4
PreSimNet (Ours)	0.258	99.23%	247,621	2.7

2. Behavior Simulation

In the closed-loop simulation task, PreSimNet achieves the lowest average RMSE for both velocity (0.342 m/s) and gap (0.420 m).

Model	Velocity RMSE (m/s)	Gap RMSE (m)
	Avg.	Avg.
Personalized IDM	0.359	1.352
PIDL-IDM (Joint)	1.126	1.085
PILSTM-IDM (Joint)	0.606	0.776
PIT-IDM (Joint)	0.855	0.677
PreSimNet (Ours)	0.342	0.420

3. Long-term Stability: Prediction vs. Simulation

Long-term (50s+) iterative rollouts highlight the synergistic advantage of our framework.

Figure: In four different car-following scenarios, conventional open-loop prediction (pink line) quickly leads to unrealistic or colliding trajectories due to error accumulation. In contrast, PreSimNet's closed-loop simulation module (yellow line) remains stable and closely follows the ground truth (blue line) over extended horizons.

---