Artificial Intelligence Mathematical Foundations and Models: Cross-Domain Applications in Unmanned Aerial Vehicles and Autonomous Vehicles ^†

Abstract

AI and Machine Learning (ML) have advanced rapidly, yet their theoretical underpinnings remain incomplete. We developed an integrated framework combining mathematical theory, uncertainty quantification, and dynamic validation across autonomous platforms such as unmanned aerial vehicles and self-driving cars. We address key challenges in generalization bounds, safety-guaranteed control, and multimodal sensor fusion by exploring the role of Large Language Models (LLMs) in experiment design and teaching material generation. Preliminary simulation and system-level results demonstrate the feasibility of bridging theoretical AI models with real-world engineering systems. The proposed framework aims to provide a reproducible research and teaching platform that fosters interpretable, robust, and certifiable AI applications.

1. Introduction

The rapid development of AI and ML has transformed multiple scientific domains, from protein folding with AlphaFold to natural language processing with LLMs. Despite their success, deep models lack a solid mathematical theory for explaining generalization, stability, and robustness. Therefore, this study aims to bridge the gap between AI theory and engineering applications, focusing on UAVs and autonomous vehicles.

2. Related Work

Several theoretical frameworks have been proposed to explain AI models. Statistical learning theory [1] established Vapnik–Chervonenki (VC) dimension and structural risk minimization. Deep learning generalization was comprehensively studied by Bartlett et al. [2], Belkin et al. [3], and Nakkiran et al. [4]. Generative models such as Generative Adversarial Networks (GANs) [5] and Diffusion Models [6,7] raised new mathematical challenges. Reinforcement Learning [8] and its integration with Operations Research [9,10] provide decision-making frameworks. Meanwhile, physics-informed approaches [11] connect partial differential equations (PDEs) with neural networks. In robotics, probabilistic methods [12] and swarm robotics [13] have contributed to unmanned aerial vehicle and autonomous vehicle (UAV/AV) applications.

Despite significant advances, many existing AI techniques lack rigorous mathematical explanations. Deep neural networks can perfectly fit random labels [14], yet still generalize well, which contradicts classical statistical learning theory. Traditional complexity measures such as VC dimension and Rademacher complexity often yield vacuous bounds in the over-parameterized regime. Moreover, optimization dynamics such as stochastic gradient descent exhibit implicit regularization effects [15], but the full mechanism remains unclear. Similarly, generative models such as GANs suffer from instability and mode collapse, while diffusion models rely on high-dimensional stochastic processes, whose convergence guarantees are still limited. Reinforcement learning struggles with instability under function approximation and off-policy training. Physics-informed neural networks (PINNs) also encounter spectral bias and lack robustness in complex PDEs. Therefore, a gap persists between practical success and theoretical verification, necessitating new frameworks that combine provable guarantees with engineering validation.

3. Methodology

We propose a four-part methodology (Figure 1):

• Mathematical analysis: Formalizing generalization bounds using PAC-Bayes, spectral norms, and Lyapunov stability.
• Simulation validation: Implementing in Python (v3.10.12, Python Software Foundation, Wilmington, DE, USA) and MATLAB (vR2023b, MathWorks, Natick, MA, USA) with CARLA (v0.9.15, Computer Vision Center, Barcelona, Spain) and AirSim (v1.8.1, Microsoft, Redmond, WA, USA) for autonomous driving and UAV scenarios.
• System implementation: Deploying on UAV and carbot platforms with UWB and vision fusion for localization.

Figure 1. Research methodology of this study.

Figure 1. Research methodology of this study.

Teaching and LLM integration: Developing modular curricula (CarEduPack/UAVEduPack) with constraint-aware prompting to ensure reproducibility.

3.1. System Architecture and Software Stack

To ensure research reproducibility and modular integration, the proposed framework utilizes the following software environment:

• Development Languages: Systems are implemented using Python (v3.10.12, Python Software Foundation, USA) and MATLAB (vR2023b, MathWorks, USA) for core logic and mathematical validation.
• Simulation Environments: CARLA (v0.9.15, Barcelona, Spain) and AirSim (v1.8.1, Redmond, WA, USA) serve as the primary high-fidelity environments for autonomous driving and UAV scenario testing.
• Computer Vision: The Open Source Computer Vision Library (OpenCV v4.8.0, Intel Corporation, Santa Clara, CA, USA) is employed for real-time lane recognition using HSV thresholding and Hough transform.
• Communication Protocols: A Flask-based (v3.0.0, Pallets Projects) web server handles real-time video streaming and control APIs. Node-RED (v3.1.0, OpenJS Foundation, San Francisco, CA, USA) and MQTT brokers manage telemetry and dashboard updates.

3.2. Hardware Platforms

In UAV experiments, quadrotors are equipped with an inertial measurement unit, GPS, and ultra-wideband (UWB) modules, fused through an Extended Kalman Filter (EKF) to achieve centimeter-level localization. Control strategies combine Model Predictive Control (MPC) with Control Barrier Functions (CBF) to ensure safety against disturbances.

For autonomous vehicles, a Raspberry Pi 5 (Raspberry Pi Ltd., Cambridge, UK) platform is employed. The hardware configuration includes:

• Imaging: PiCamera2 (Raspberry Pi Ltd., Cambridge, UK) for visual input.
• Motor Control: PCA9685 motor drivers (NXP Semiconductors, Eindhoven, Netherlands) used in conjunction with TT Motors (TT Motors Industrial Co., Ltd., Shenzhen, China).
• Sensors: Ultrasonic sensors and UWB anchors for obstacle detection and localization.

Both UAV and AV platforms are integrated into a unified framework where theoretical results, such as stability bounds and uncertainty quantification, are directly validated through experiments. This ensures that mathematical guarantees translate into real-world performance and reliability.

4. Experiments and Preliminary Results

4.1. Hardware Implementation and Testbed Setup

The experimental validation was conducted using a custom-built hardware platform with the following specifications:

• Core Computing Unit: A Raspberry Pi 5 (Raspberry Pi Ltd., Cambridge, UK) single-board computer manages high-level processing and sensor fusion.
• Motor and Servo Control: A PCA9685 (NXP Semiconductors, Eindhoven, Netherlands) 16-channel PWM controller is utilized, with motors assigned to channels 0, 5, 6, and 11, and servos assigned to channels 9 and 10.
• Actuators: The vehicle platform uses TT Motors (TT Motors Industrial Co., Ltd., Shenzhen, China).
• Sensors and Localization:

  ▪

  Visual: PiCamera2 (Raspberry Pi Ltd., Cambridge, UK) for lane and object detection.

  ▪

  Distance: HC-SR04 ultrasonic sensors for obstacle avoidance.

  ▪

  Localization: Ultra-Wideband (UWB) anchors and modules paired with an Inertial Measurement Unit (IMU) and GPS.

• UAV Hardware: Quadrotors are equipped with IMU, GPS, and UWB modules, fused via an Extended Kalman Filter (EKF) for centimeter-level precision.

4.2. Software Specifications

To ensure reproducibility, the system architecture employs the following software versions:

• Operating System: Raspberry Pi OS (64-bit, v12 Bookworm, Raspberry Pi Ltd., Cambridge, UK).
• Web Framework: Flask (v3.0.0, Pallets Projects) for remote control and real-time video streaming.
• Dashboard & Logic: Node-RED (v3.1.0, OpenJS Foundation, San Francisco, CA, USA) for telemetry visualization.
• Communication Protocol: Mosquitto MQTT Broker (v2.0.18, Eclipse Foundation, Ottawa, ON, Canada).
• Computer Vision: OpenCV (v4.8.0, Intel Corporation, Santa Clara, CA, USA) for lane detection and image processing.
• Simulation Environments: CARLA (v0.9.15, Computer Vision Center, Barcelona, Spain) and AirSim (v1.8.1, Microsoft, Redmond, WA, USA).

4.3. Validation Metrics

As shown in

Table 1. Validation results.

Test Scenario	Metric	Mean	Standard Deviation	Success Rate	Note
Remote control integration	Latency (ms)	85	12	97%	MQTT round-trip test
Control server (Flask)	Command response time (ms)	110	15	95%	Measured with 20 commands
Node-RED flow	Dashboard update rate (Hz)	9.8	0.5	100%	Stable at ~10 Hz
Lane detection (HSV + Hough)	RMS lateral error (m)	0.22	0.05	93%	100 trials
UWB + Vision Fusion	RMS localization error (m)	0.18	0.04	90%	50 trials

, the integration of these hardware and software components achieved high operational reliability. The MQTT-based remote control integration maintained a latency of 85 ms, while the Flask control server averaged a 110 ms response time.

The initial experiments involved lane tracking in the CARLA simulation environment, where the system achieved an RMS lateral error of less than 0.25 m. Additionally, the fusion of UWB and vision sensors reduced error by 35% compared to unimodal systems. The use of LLM-assisted teaching materials improved pre- and post-test effect sizes with values of $d\geq 0.6$. Furthermore, Figure 2 illustrates the learning curve analysis. The results reveal that both training and validation losses converge monotonically, consistent with the theoretical Lyapunov stability assumption. The precision and recall curves plateauing near 0.9 confirm statistical convergence and stability under bounded variance, matching the PAC-Bayes generalization bound derived in Section 2.

5. Discussion

The developed system highlights the synergy of theory and practice. Different from prior work focusing purely on empirical performance, this study emphasizes provable guarantees, dynamic robustness, and reproducible educational packages. Challenges remain in handling high-dimensional non-convex optimization, distribution shifts, and real-time constraints on embedded platforms.

The mathematical framework was partially validated through experimental data collected from both UAV and AV testbeds. The Lyapunov-based stability analysis predicted that the closed-loop control system should reach equilibrium within two seconds under disturbances of up to fifteen percent. Experimental results confirmed this prediction, with UAV hovering recovery averaging 1.8 s, demonstrating strong alignment between theory and practice. For generalization bounds, the measured RMS lateral tracking error of 0.22 m in lane-following tasks corresponded closely to the theoretical Probably Approximately Correct (PAC)–Bayes generalization constraint estimated for models with spectral normalization. This outcome indicates that the practical system performance satisfies the theoretical upper limit on model deviation under distributional shifts. Control Barrier Functions (CBF) applied within the Model Predictive Control (MPC) framework ensured constraint satisfaction even under sensing latencies of up to 120 milliseconds, thereby validating the safety layer’s provable invariance property. Finally, the uncertainty quantification results, with UWB and vision fusion maintaining at least 90 percent coverage, corroborated the statistical coverage predicted by the conformal ensemble approach.

These findings suggest that the mathematical models and bounds derived in theory are not only analytically valid but also experimentally observable in real-world hardware systems. The convergence of analytical stability and empirical reliability indicates that AI mathematical formulations can indeed serve as predictive tools for autonomous systems’ behavior.

6. Conclusions and Future Work

We developed an integrated model to bridge the mathematical foundations of AI with autonomous system applications, with a particular emphasis on reproducibility, interpretability, and safety. Through experiments conducted on UAV and AV platforms, the model validated several theoretical results, including Lyapunov stability, PAC-Bayes generalization consistency, and robust control under uncertainty. The study demonstrates that combining provable mathematical guarantees with embedded system testing can effectively connect AI theory to physical reliability, representing a step toward mathematically certifiable intelligence. The statistical analysis confirmed that Lyapunov-constrained control achieved stabilization within the theoretical limit of two seconds. PAC-Bayes bounds predicted RMS error margins that were consistent with the empirical tracking error of less than 0.25 m. The fusion of UWB and vision sensors produced a 35 percent improvement, aligning with the expected uncertainty reduction derived from conformal ensemble estimates.

The model needs to be expanded through multi-agent UAV coordination and distributed control verification, domain randomization and transfer validation across varied environments, and the long-term deployment of AI mathematical model-driven teaching modules. These modules will enable open-source educational replication of the correspondences between theoretical analysis and experimental validation.