PCARNN-DCBF Enables Real-Time Geofence Enforcement for Ground Vehicles with Control-Affine Residual Neural Networks

Runtime geofencing, a crucial technology for defining and enforcing the safe operating limits of ground vehicles, currently faces challenges in balancing accurate performance with the need for provably safe control. Yinan Yu and Samuel Scheidegger, from Asymptotic AI, tackle this problem by presenting PCARNN-DCBF, a new system that combines a physics-informed neural network with a safety-critical control strategy. This innovative pipeline explicitly incorporates the fundamental principles of vehicle movement, ensuring the necessary conditions for reliable optimisation and safe operation, unlike many existing approaches. The team demonstrates that this structure-preserving method significantly improves performance and safety in complex driving scenarios, outperforming both traditional analytical techniques and less structured neural network models.

Vehicle Dynamics Control and Prediction Models

This research details experimental results for various control and prediction models applied to vehicle dynamics within a simulation environment. The experiments focus on controlling and predicting vehicle behavior, with a focus on stability and trajectory tracking. Two different vehicle models, Audi and Lincoln, were used throughout the experiments. The team evaluated several models, including PCARNN, which combines recurrent neural networks for time-series prediction with probabilistic control techniques, and residual networks, a type of deep neural network known for learning complex patterns. Neural Ordinary Differential Equations, a more advanced type of neural network modeling continuous-time dynamics, also showed promising results, while a simplified bicycle model served as a baseline for comparison.

Performance was assessed using metrics including Control Failure, a binary indicator of system instability, and Root Mean Squared Error, which measures the accuracy of vehicle state predictions. The value of a parameter, gamma, significantly affected control performance, with higher values potentially leading to more aggressive, but potentially unstable, control actions. The Audi vehicle generally performed better than the Lincoln vehicle, potentially due to differences in vehicle dynamics or control algorithms. The research suggests that PCARNN is a viable approach for controlling and predicting vehicle dynamics, but careful hyperparameter tuning is essential to achieve optimal performance. Developing vehicle-specific models tailored to the unique dynamics of each vehicle may prove beneficial, and results from the simulation environment should be validated in a real-world setting to ensure generalizability.

Physics-Preserving Neural Networks for Safe Geofencing

This study pioneers a novel approach to runtime geofencing for ground vehicles, integrating a Physics-encoded Control-Affine Residual Neural Network with a preview-based Discrete Control Barrier Function, termed PCARNN-DCBF. This system addresses the challenge of reconciling high-fidelity learning with the structural requirements necessary for verifiable control, a critical aspect of Operational Design Domains. Unlike conventional learned models, PCARNN explicitly preserves the control-affine structure of vehicle dynamics, ensuring the linearity required for reliable optimization and predictive control. The core of the method involves training the residual neural network to accurately model vehicle dynamics while maintaining the inherent control-affine structure.

This network then provides predictions used within the DCBF, which formulates the geofencing problem as a constrained optimization solved via a real-time Quadratic Program. Experiments were conducted within the CARLA simulation environment, utilizing both electric and combustion vehicle platforms to demonstrate the system’s versatility. The study’s innovative approach extends beyond simple boundary detection, focusing on predictive intervention. The preview-based DCBF anticipates potential violations by considering future vehicle states, enabling proactive control adjustments. This predictive capability is crucial for maintaining safety and stability, particularly in dynamic environments, and the reliance on a Quadratic Program ensures computational efficiency for real-time operation. The results demonstrate a significant performance improvement over existing methods, highlighting the effectiveness of the structure-preserving approach in achieving robust and verifiable runtime geofencing.

Real-Time Geofencing with Neural Networks and Control

Scientists developed a novel pipeline, PCARNN-DCBF, to improve runtime geofencing for ground vehicles, a critical component of Operational Design Domains. This work addresses the challenge of reconciling high-fidelity learning with the structural requirements needed for verifiable control systems by integrating a Physics-encoded Control-Affine Residual Neural Network with a preview-based Discrete Control Barrier Function. Experiments demonstrate that PCARNN-DCBF explicitly preserves the control-affine structure of vehicle dynamics, a key feature enabling the DCBF to enforce polygonal keep-in constraints via a real-time Quadratic Program. This program effectively handles high relative degree and mitigates actuator saturation, crucial for precise vehicle control.

The team measured performance across both electric and combustion vehicle platforms in challenging driving regimes, consistently outperforming analytical and unstructured neural baselines. The researchers utilized a dynamic bicycle model to accurately capture lateral and longitudinal behavior, and the neural network component of the pipeline was trained using both data and physics-based regularization, achieving a robust and accurate prediction of vehicle state derivatives. Measurements confirm the system’s ability to maintain vehicle position within defined spatial constraints, demonstrating a significant advancement in geofencing technology and delivering improved safety and reliability for autonomous vehicle operation in complex environments.

Structure-Preserving Geofencing For Vehicle Safety

This research presents a novel geofencing framework, PCARNN-DCBF, designed to enhance the safety of ground vehicles operating within defined boundaries. The team successfully integrated a physics-encoded control-affine residual neural network with a preview-based discrete control barrier function, creating a system that combines the strengths of both learned models and analytical control techniques. By explicitly preserving the control-affine structure of vehicle dynamics within the learning process, the framework ensures the safety constraint remains linear, enabling the use of a fast and interpretable optimization process for real-time intervention. Comparative analysis demonstrated that this structure-preserving approach significantly outperforms both purely analytical methods and unstructured neural baselines in maintaining containment reliability. The research also revealed that drivetrain topology influences optimal learning architecture, with decoupled electric powertrains benefiting from split network designs and coupled combustion systems requiring shared representations.