|
Our journal welcomes not only original high-quality papers covering the theoretical, experimental and operational aspects of electrical and electronics engineering in mobile radio, motor vehicles and land transportation, but also industry-focused publication focusing on research findings and suggesting ideas that may be useful to those conducting similar research.
Below, we highlight two featured peer-reviewed articles:
Our first monthly feature paper, written by authors from the University of New Mexico, presents a clustered federated reinforcement learning framework that enables UAVs to collaboratively train navigation models across distributed environments. By aggregating local experiences within clusters rather than relying on centralized data collection, the approach reduces communication overhead and improves real-world adaptability of models initially trained in simulation, achieving effective control performance under complex and uncertain conditions.
Complementing this is another feature paper co-authored by University of Galway and Valeo Vision Systems, which introduces a lightweight perception method for detecting unseen or anomalous objects in autonomous driving systems. Using multi-camera inputs and a new NuScenesOOD dataset, the authors identifies out-of-distribution objects in the Bird’s-Eye View space, enhancing the vehicle’s ability to detect novel or irregular obstacles.
We’ve provided short summaries of these feature articles, written in accessible language that we hope will make your reading experience enjoyable.
Clustered Federated Reinforcement Learning for Autonomous UAV Control in Air Corridors
Meng Xiang Xuan, Liangkun Yu, Xiang Sun, and Sudharman K. Jayaweera
Summary by Meng Xiang Xuan: Reinforcement learning shows promise in 3D maneuvering tasks, but it requires massive amounts of data for training, and its performance decreases when the training environment differs from the evaluation one, such as when transitioning from a simulation to reality. If using real data to train “on the fly”, transmitting and processing the amount of data required for a single model update requires computing power and energy that may not be available (for example: on a quadcopter or an interplanetary probe).
Federated learning can help mitigate this by only transmitting and averaging model parameters, which are orders of magnitude less bulky than the data required for an equivalent update. Unlike traditional federated reinforcement learning (FRL) frameworks that assume every client has a pre-existing local dataset, our method organizes its clients into clusters, where each cluster head aggregates flight data from its members to perform local training before contributing to a global model. For example, multiple UAV test sites could share their training experience in this manner, each cluster headed by a local server capable of backpropagation.
We use our method to adapt a model trained in simulation to an uncertain environment, where multiple entities traverse a series of air corridors (straight cylinders and curved tori) without colliding. We compare the performance to a traditional backpropagation method and LoRA, a training method that reduces the number of weights trained. Effectively, the traditional method and LoRA are represented by a single agent collecting all the data, while our method has three agents each collecting a third of the total data. While our method slightly underperforms compared to the traditional method (89% versus 95% arrival rate after 10 sequential corridors), it shows clear improvement over the untrained method, and retains some performance even if the number of simultaneous entities in the corridors is increased. Further evaluation is required on clustering parameters (number of clusters, clients per cluster) and the effect of heterogeneous clusters on the global model.
Full article: IEEE Open Journal of Vehicular Technology, Volume 6
Towards Robust Autonomous Driving: Out-of-Distribution Object Detection in Bird's Eye View Space
Muhammad Asad, Ihsan Ullah, Ganesh Sistu, Michael G. Madden
Summary by Muhammad Asad: Supervised self-driving perception algorithms are good at handling common types of objects such as cars, buses, signals and pedestrians. But in real environment, there can be arbitrary objects that were not represented in the training data. These are called out-of-distribution, or OOD, objects. Our work introduces an approach to detect these objects in the car’s top-down view, called the Bird’s-Eye View (BEV).
Our method is lightweight and easy to add to current BEV systems. It uses inputs from multiple cameras and generates an OOD map. In training, we introduce random patches on some vehicles that completely hide the vehicle on the drivable area grid. Each object in the map gets a score showing how different it is from the learned mean distribution of known objects, and a high score means the object is OOD. This also helps the system to avoid high-confidence but incorrect classifications when it sees something new, for example, animals on the road, random debris, or unusual objects. Even if it cannot name these objects, it can still find where are they located.
We also introduce a new dataset called NuScenesOOD. Here we add different patterns and augmentations on the vehicles to make them unusual and different from other vehicles. We also use this dataset for evaluation purposes.
Our method targets vehicle-shaped anomalies in the BEV Space, maintaining high accuracy for known and enhancing detection of unknown objects. This research contributes to making future autonomous vehicles safer by improving their ability to detect diverse vehicle like OOD objects in their environment.
Full article: IEEE Open Journal of Vehicular Technology, Volume 6
About the IEEE Open Journal of Vehicular Technology (OJVT)
The IEEE OJVT covers the theoretical, experimental and operational aspects of electrical and electronics engineering in mobile radio, motor vehicles and land transportation. A brief summary of these fields of interest are as follows:
- Mobile radio shall include all terrestrial mobile services
- Motor vehicles shall include the components and systems and motive power for propulsion and auxiliary functions
- Land transportation shall include the components and systems used in both automated and non-automated facets of ground transport technology
| 
