Artificial Drivers: the new frontier of autonomous driving on road and track

Over the past five years, autonomous vehicle competitions have become real experimental labs for smart mobility. Prestigious events such as the Abu Dhabi Autonomous Racing League and Formula Student Driverless have brought Formula 1-style cars to race without a human driver, pushing real-time planning, control, and perception capabilities to the limit. However, these challenges have also highlighted unresolved obstacles that arise when a vehicle must operate at high speeds and handle near-limit driving conditions.

In this context, the development of advanced mathematical models is crucial for trajectory planning and real-time vehicle control. These models must strike a balance between computational efficiency—vital for quickly responding to unforeseen scenarios—and robustness, to ensure safety in complex dynamic conditions. Research at the Department of Industrial Engineering focuses on three key unresolved issues in both academic literature and current applications:

1. Developing efficient models for real-time minimum-time trajectory planning and vehicle control, accurately capturing vehicle dynamics.
2. Executing safe and optimal trajectories in the presence of dynamic opponent vehicles.
3. Implementing vehicle dynamics learning methods that require little experimental data and can generalize to previously unseen driving conditions.
   

An Artificial Driver for Autonomous Racing

One of the main outcomes of the research is the development of an artificial driver for autonomous track racing, capable of computing and following minimum-time optimized trajectories. At the core of this system is an innovative architecture that combines physics-inspired neural networks with optimization methods based on Newton’s laws to describe vehicle mechanics. This approach ensures a higher level of generalization than traditional machine learning techniques, enabling the artificial driver to effectively adapt to unknown tracks and race conditions. Tested in advanced simulations and on 1:8 scale real vehicles, the system has shown superior performance in handling complex scenarios and tracks never explored during training.

Trajectory Planning to Overtake Opponent Vehicles

The second research contribution focuses on the ability to avoid and overtake competing vehicles. To address this, a neural network-based trajectory planner was developed, capable of generating a minimum-time maneuver graph to prevent collisions and overtake opponents. The algorithm was successfully tested in competitive scenarios involving multiple autonomous vehicles, demonstrating its ability to predict and manage dynamic interactions with competitors without compromising overall performance.

Adaptive Learning for Autonomous Parking in Urban Scenarios

Applying the same trajectory planning and control principles, the research extends to urban contexts, with particular attention to autonomous parking in complex environments. An algorithm was developed that enables a vehicle to perform real-time optimized parking maneuvers, adapting to various spatial configurations. Simulation tests confirmed the system’s robustness, even in the presence of vehicle parameter variations and sensor noise.

Advanced Lateral Velocity Estimation for Autonomous Control

The final part of the research introduces an innovative method for estimating the vehicle’s lateral velocity—a critical variable for stability systems such as ESP and for autonomous driving near the limits of grip. The proposed method uses neural networks with architectures inspired by the vehicle’s kinematic equations, achieving greater accuracy than traditional machine learning algorithms. This solution not only requires less training data but has also proven highly effective in adapting to different driving conditions and a wide range of vehicle types.

Conclusions and Future Perspectives

The results of this research not only improve the performance of autonomous vehicles in competitive settings but also lay the groundwork for the evolution of autonomous driving on public roads, making it safer, more interpretable, and better adapted to complex scenarios. The solutions developed are already the subject of collaborations with automotive companies and will soon be tested on experimental vehicles, marking a decisive step toward a future where artificial intelligence and vehicle dynamics converge to redefine the very concept of mobility.