Panagiotis Angeloudis’s research while affiliated with Imperial College London and other places

Achieving safety in autonomous driving through Multi-Agent Reinforcement Learning (MARL) is a critical yet challenging task due to non-stationarity, partial observability, and the need for effective coordination among agents. Although earlier cooperative MARL methods have aimed to improve coordination by sharing encoded state observations, we find these strategies are insufficient in safety-critical scenarios. To address this gap, we present IntNet, a novel MARL framework that enhances safety by incorporating the transmission of vehicle intent and adaptive communication scheduling into a unified end-to-end learning paradigm, jointly optimising all components for safe coordination. Central to our approach is an observation prediction module that enables agents to forecast subsequent policy outputs, which they can then share across a vehicle-to-vehicle network. Our model-free architecture employs a Graph Attention Network to process incoming messages and a scheduler component to dynamically optimise the communication graph for bandwidth efficiency. We compare IntNet against state-of-the-art communicative MARL methods in complex urban environments with both autonomous and human-operated vehicles, achieving improved coordination, lower collision rates, and reducing communication efforts by up to 60%. Through extensive experiments, we assess the framework's learning efficiency, scalability, and the balance between information sharing and bandwidth usage for collision-free trajectories.

Traffic accidents, especially at intersections, are a major road safety concern. Previous research has extensively studied intersection-related accidents, but the effect of building-induced visibility restrictions at intersections on accident rates has been under-explored, particularly in urban contexts. Using OpenStreetMap data, the UK's geographic and accident datasets, and the UK Traffic Count Dataset, we formulated a novel approach to estimate accident risk at intersections. This method factors in the area visible to drivers, accounting for views blocked by buildings - a distinctive aspect in traffic accident analysis. Our findings reveal a notable correlation between the road visible percentage and accident frequency. In the model, the coefficient for "road visible percentage" is 1.7450, implying a strong positive relationship. Incorporating this visibility factor enhances the model's explanatory power, with increased R-square values and reduced AIC and BIC, indicating a better data fit. This study underscores the essential role of architectural layouts in road safety and suggests that urban planning strategies should consider building-induced visibility restrictions. Such consideration could be an effective approach to mitigate accident rates at intersections. This research opens up new avenues for innovative, data-driven urban planning and traffic management strategies, highlighting the importance of visibility enhancements for safer roads.

Accurately predicting traffic risks at urban intersections is essential for improving road safety. While traditional models use data sources like road traffic conditions, geometry, and signals, they often miss the spatial interactions between road networks and buildings. This study introduces a multimodal deep learning framework that integrates aerial imagery, building footprint data, and traffic flow information to improve traffic risk prediction and better capture these complex relationships. By leveraging datasets from OpenStreetMap, the UK Traffic Count, and high-resolution aerial imagery, our approach creates a comprehensive representation of the urban environment, capturing intricate spatial relationships between road networks, surrounding structures, and traffic conditions. Using DeepLabV3+, UNet++, and SegFormer as baseline models, we demonstrate that combining building and traffic data enhances prediction accuracy compared to models relying solely on visual data. Our results show that the DeepLabV3+ model, when incorporating both building and traffic data, achieves the highest Intersection over Union (IoU) score of 0.4052 and the lowest Root Mean Square Error (RMSE) of 0.0907. These findings underscore the effectiveness of a multimodal approach in traffic risk assessment, offering a more precise tool for urban planning and traffic management interventions. The code and data used in this study are available at https://github.com/zachtian/Multimodal-Learning-for-Traffic-Risk-Prediction.

Autonomous vehicles need to continuously analyse the driving context and establish a comprehensive understanding of the dynamic traffic environment. To ensure the safety and efficiency of their operations, it would be beneficial to have accurate predictions of surrounding vehicles’ future trajectories. AVs can adjust their motions proactively to improve road safety and comfort with such information. This paper proposes a novel approach to predict the future trajectories of interacting vehicles, through a model of potential spatial-temporal interactions. A unique kernel function that emphasises risk-awareness was developed to extract spatial dependencies. The established model was trained and evaluated with the publicly available Highway Drone Dataset and Intersection Drone Dataset. The performance of the developed model was assessed with eight state-of-the-art methods. An ablation study and safety analysis were also conducted to evaluate the proposed risk-awareness kernel function. Results show that the proposed model’s inference speed is over eight times faster than the commonly used LSTM-based models. It also achieves an improvement of over 8% in prediction accuracy when compared with the state-of-the-art model.

Collaborative navigation becomes essential in situations of occluded scenarios in autonomous driving where independent driving policies are likely to lead to collisions. One promising approach to address this issue is through the use of Vehicle-to-Vehicle (V2V) networks that allow for the sharing of perception information with nearby agents, preventing catastrophic accidents. In this article, we propose a collaborative control method based on a V2V network for sharing compressed LiDAR features and employing Proximal Policy Optimisation to train safe and efficient navigation policies. Unlike previous approaches that rely on expert data (behaviour cloning), our proposed approach learns the multi-agent policies directly from experience in the occluded environment, while effectively meeting bandwidth limitations. The proposed method first prepossesses LiDAR point cloud data to obtain meaningful features through a convolutional neural network and then shares them with nearby CAVs to alert for potentially dangerous situations. To evaluate the proposed method, we developed an occluded intersection gym environment based on the CARLA autonomous driving simulator, allowing real-time data sharing among agents. Our experimental results demonstrate the consistent superiority of our collaborative control method over an independent reinforcement learning method and a cooperative early fusion method.

With growing air travel demand, weather disruptions cost millions in flight delays and cancellations. Current resilience analysis research has been focused on airports and airlines, rather than the en-route waypoints, and has failed to consider the impact of disruption scenarios. This paper analyses the resilience of the United Kingdom (UK) air traffic network to weather events that disrupt the network’s high-traffic areas. A Demand and Capacity Balancing (DCB) model is used to simulate adverse weather and re-optimise the cancellation, delay, and rerouting of flights. The model’s feasibility and effectiveness were evaluated under 20 concentrated and randomly occurring extreme disruption scenarios, lasting 2 h and 4 h. The results show that the network is vulnerable to extended weather events that target the network’s most central waypoints. However, the network demonstrates resilience to weather disruptions lasting up to two hours, maintaining operational status without any flight cancellations. As the scale of disruption increases, the network’s resilience decreases. Notably, there exists a threshold beyond which further escalation in disruption scale does not significantly impair the network’s performance.

Wind speed affects aviation performance, clean energy production, and other applications. By accurately predicting wind speed, operational delays and accidents can be avoided, while the efficiency of wind energy production can also be increased. This paper initially overviews the definition, characteristics, sensors capable of measuring the feature, and the relationship between this feature and wind speed for all Quality Indicators (QIs). Subsequently, the feature importance of each QI relevant to wind-speed prediction is assessed, and all QIs are employed to predict horizontal wind speed. In addition, we conduct a comparison between the performance of traditional point-wise machine learning models and temporally correlated deep learning ones. The results demonstrate that the Bidirectional Long Short-Term Memory (BiLSTM) neural network yielded the highest level of accuracy across three metrics. Additionally, the newly proposed set of QIs outperformed the previously utilised QIs to a significant degree.