This repository is the replication package for COIN: Collaborative Interaction-Aware Multi-Agent Reinforcement Learning for Self-Driving Systems. It provides the codebase, dependencies, and instructions needed to reproduce the training and evaluation results of COIN in multi-agent self-driving scenarios.

ABSTRACT: Advanced driver assistance systems (ADASs) can greatly enhance road safety by providing real-time warnings to drivers in imminent crash situations. However, the provided warning time may deviate from its designed time. There is limited research on how warning uncertainties influence drivers’ behavior, safety performance, and trust. This study conducted a driving simulator study to examine how uncertainties in warnings impact driving behaviors and trust using a roundabout driving scenario. Two warning error distributions were constructed to represent low and high warning uncertainty levels. Thirty-six participants were recruited and randomly divided into two groups under the two uncertainty levels in a driving simulator experiment. The betweengroup analysis shows that the lower warning uncertainty level group results in higher trust and that trust increases (or decreases) over time under low (or high) uncertainty levels. The within-group analysis shows that higher warning errors downgrade drivers’ trust and safety performance when the errors are high. 

Heavy-haul railways are critical for transporting freight. However, prolonged wheel–rail interactions cause frequent rail defects, particularly in small-radius curve sections. Ultrasonic A-scan signals are essential for the non-destructive evaluation of internal rail defects. In real heavy-haul environments, these signals suffer from strong non-Gaussian coupled noise. Such noise includes structural noise, low-frequency irrelevant components, and high-frequency electrical noise. Noise aliasing obscures defect echoes and increases the risk of missed detections. Conventional denoising methods are limited by poor noise–signal separability, mode mixing, and inadequate adaptability to complex non-Gaussian signals. To address these challenges, an A-scan signal model under noise-coupled conditions is constructed by analyzing the statistical and time–frequency characteristics of different noise components.

The replication package contains the questionnaire, datasets and analysis process used for the study.

The basic code and date  of DLE

STPredictor is an explainable ship trajectory prediction framework powered by large language models. It reformulates trajectory prediction as a language modeling task using natural-language prompts and integrates Chain-of-Thought reasoning for more transparent and reliable predictions. Experiments on two large-scale AIS datasets show strong performance and improved interpretability over existing baselines.

This replication package provides the complete source code and configuration files used in the study “CogDrive: Cognition-Driven Multimodal Prediction-Planning Fusion for Safe Autonomy.”

The package enables full reproduction of the proposed cognition-driven multimodal trajectory prediction and safety-aware planning framework, including model architecture implementation, training procedures, evaluation scripts, and planning modules.

The provided code supports both open-loop prediction evaluation and closed-loop planning simulations as reported in the manuscript. All scripts are organized to facilitate reproducibility of the main experimental results on public autonomous driving benchmarks.

This paper identifies, analyzes, and addresses case-level forgetting in continual learning for trajectory prediction. We propose the Dynamically Expandable Interactive Trajectory Predictor (DEITP), a novel framework that preserves previously learned knowledge through a dynamic model expansion mechanism. The mechanism regulates expansion timing by assessing model similarity, thereby controlling model growth while preventing catastrophic forgetting. Furthermore, to operate in realistic task-free settings where task identity is unavailable at test time, we introduce a task identification strategy based on a familiarity autoencoder that selects the most appropriate expert for prediction. Extensive experiments on real-world datasets demonstrate that DEITP substantially mitigates forgetting and achieves zero-forgetting performance when task identities are known.

The entire project package is uploaded. Simply unzip it and run the train and test functions separately; the data have already been preprocessed.

Two standard datasets on travel mode choice are used, namely LPMC and Swissmetro (SM). The LPMC dataset consists of single-day travel diary data obtained by the London Travel Demand Survey from 2012 to 2015. The dataset includes 81,096 samples, each of which corresponds to a trip taken by a person in one of the 17,616 households participating in the survey. The dataset contains four travel modes, namely walking, cycling, public transportation (PT), and car. Moreover, the SM dataset comprises passenger survey data gathered in Switzerland. It includes samples from 1291 passengers across nine travel scenarios designed with three kinds of modes, i.e., train, SM, and car. In the materials, we make separate divisions based on the datasets. Among them, the "data" file contains the original data, the processed data, and the data processing code. The configuration details are described in the "conf" folder. The analysis code is included in the "utilities" file. Then, the codes for the centralized, federated learning framework and the federated meta-learning framework are also attached.

This paper proposes AdvGLOW, a framework generating covert adversarial attacks against autonomous driving perception systems. It uses a Glow-based reversible neural network for bi-directional image-latent space transformation, crafting imperceptible perturbations. Coupling layers and actnorm ensure invertible, effective transformations. The method achieves real-time generation (<50 ms). Experiments on driving datasets and models confirm its stealth and effectiveness, pioneering the use of normalizing flows for physically realizable attacks in this context. Repository includes code and configurations.

Mandatory lane changes pose significant challenges to trajectory planning at intersections, where vehicles are required to change lanes mid-block to reach designated turn lanes before the stop bar. MLCs often generate shockwaves that induce increased vehicle delay and fuel consumption, and the presence of human-driven vehicles in mixed traffic further exacerbates this issue. To address these challenges, this study formulates the joint longitudinal-lateral trajectory planning problem in mixed traffic as a multi-agent reinforcement learning task. We propose SS-MA-PPO, a Simulation-Supervised Multi Agent Proximal Policy Optimization framework, which guides connected and automated vehicles in both acceleration and lane-change decisions. A Simulation-Guided Supervisory Module performs offline trajectory rollouts of human-driver models to assess feasibility and safety, and arbitrates online between rule-based and learned policies. The information of surrounding vehicles is incorporated in the observation to achieve vehicle cooperation, and a transfer learning mechanism is designed to accelerate training.

This replication package contains the dataset (CTSU) and source code used in TrafficPerceiver.

All the materials source coding and results coding of  Vehicle-Infrastructure Cooperative General Object Detection through Feature Flow and Differentiable Pose-based Spatial Alignment

Accurate short-horizon trajectory prediction is crucial for safe and reliable autonomous driving. However, existing vision-language models (VLMs) often fail to accurately understand driving scenes and generate trustworthy trajectories. To address this challenge, this paper introduces KEPT, a knowledge-enhanced VLM framework that predicts ego trajectories directly from consecutive front-view driving frames. KEPT integrates a temporal frequency–spatial fusion (TFSF) video encoder, which is trained via self-supervised learning with hard-negative mining, with a k-means & HNSW retrieval-augmented generation (RAG) pipeline. Retrieved prior knowledge is added into chain-of-thought (CoT) prompts with explicit planning constraints, while a triple-stage fine-tuning paradigm aligns the VLM backbone to enhance spatial perception and trajectory prediction capabilities. This replication package includes all materials required for readers to understand and reproduce the analyses reported in the paper.

This paper introduces latency-aware trajectory prediction, a new task that explicitly accounts for latency and repurposes it as a useful signal. We present LatenAux, a consolidated auxiliary learning paradigm that first decouples prediction into two tasks: a primary task that predicts valid-horizon trajectories from historical data, and an auxiliary task that utilizes latency-inclusive observations. By allowing the auxiliary branch access to latency-crafted inputs, LatenAux is then committed to transferring latency-aware knowledge to the primary branch via a progressive feature alignment strategy. Extensive experiments on two large-scale real-world datasets demonstrate the effectiveness and superiority of LatenAux, showing that it consistently supports latency-aware modeling and delivers more accurate and reliable trajectory forecasts.

This paper proposes a lightweight vision and language large model based approach for pedestrian crossing intention prediction.We introduce a two-stage tuning strategy designed to enhance the model’s explicit and implicit reasoning capabilities. In our experiments, we first use Gemini to generate chain-of-thought annotations from the raw data. We then fine-tune a lightweight Qwen 2.5 model on the resulting CoT dataset using the proposed two-stage procedure. Experiments are conducted on two public datasets, PIE and JAAD. This replication package includes all materials required for readers to understand and reproduce the analyses reported in the paper.

• .mat file DLC_acc_FAC_RAC – DLC acceleration test with Front & Rear Active Camber actuators active
• .mat file DLC_acc_FAC_RWS – DLC acceleration test with Front Active Camber & Rear Wheel Steering actuators active
• .mat file DLC_acc_FAC_RAC_RWS – DLC acceleration test with Front & Rear Active Camber and Rear Wheel Steering actuators active
• .mat file PoFF_FAC_RAC – Power-off reaction test with Front & Rear Active Camber actuators active
• .mat file PoFF_FAC_RWS – Power-off reaction test with Front Active Camber & Rear Wheel Steering actuators active
• .mat file PoFF_FAC_RAC_RWS – Power-off reaction test with Front & Rear Active Camber and Rear Wheel Steering actuators active
• Script file DLC_acc – used to generate Figure 8 of the manuscript
• Script file PoFF – used to generate Figure 9 of the manuscript

There are three parts in the replication package:

• Nowcasting

The Nowcasting folder contains the complete pipeline to train the CNN model (Python Code) with images of different vehicle classes and subsequently use the trained model to classify vehicles from video data and estimate their emissions. 

• ODME

The ODME folder includes the software DTALite as well as its input data for estimating OD demand. 

• Simulation

The simulation folder contains the raw demand data extracted from Dynameq initially calibrated by City of Stockholm for the entire Stockholm area and the scripts to transform the demand first into MATSim format and subsequently into SUMO format to realize the hybrid simulation format. Ultimately following this pipeline, the emission for the study area in Kista can be realized for given demand of the current scenario and alternative future scenarios.

To address the limited extensibility of standardized message format this study proposes a modular, edge-intelligent framework — The mobility Operating System (mOS) — integrated with a mixed-reality testbed for realistic validation of infrastructure-guided autonomous vehicle coordination. We analyzed to verify the feasibility of the C-V2X Framework for autonomous vehicle guidance in real-time at the physical testbed. The dataset was collected from the testbed experiment to analyze framework performance (speed profiles, post-encroachment time, and numerical error) and service performance (latency, jitter, and packet loss).