A Certi􀀂able AI-Based Braking Control Framework for Landing Using Scienti􀀂c Machine Learning

Aug 5, 2026

Upload by Asher

This paper proposes an AI-based aircraft braking control system for landing. Using scientific machine learning, an agent is trained to select braking strategies that adapt to landing speed and environmental conditions while enforcing physically consistent behavior by embedding landing-physics principles into the algorithm. Results show successful deceleration without skidding across runway conditions and landing speeds. The method preserves performance and safety under brake degradation and initial yaw-angle perturbations, supporting certification needs for safety-critical AI systems.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Introduction

Background and limitations of rule-based landing braking control

Prior studies and trends in anti-skid and learning-based control

Scientific machine learning motivation

AI certification and regulatory frameworks

2024 AIAA DATC/IEEE 43rd Digital Avionics Systems Conference (DASC), 29 September-3 October 2024, San Diego, CA, USA DOI:10.1109/DASC62030.2024.10749078 A Certi􀀂able AI-Based Braking Control Framework for Landing Using Scienti􀀂c Machine Learning Mevlut Uzun, Ugurcan Celik, Guney Guner, Orhan Ozdemir, Gokhan Inalhan Autonomy and AI, Cran􀀂eld University Cran􀀂eld, United Kingdom Email : {mevlut.uzun, ugurcan.celik, guney.guner, orhan.ozdemir, inalhan}@cran􀀂eld.ac.uk Abstract—This paper proposes an AI-based braking control system for aircraft during landing. Utilizing scienti􀀂c machine learning, we train an agent to apply the most effective braking strategy under various landing conditions. This approach ensures physically consistent outputs by grounding the algorithm in the principles of landing physics. Our results demonstrate that the aircraft can successfully decelerate without skidding across all runway conditions and landing speeds. Additionally, the algorithm maintains performance and safety even when brake performance degradation and initial yaw angles are introduced. This robustness is crucial for the certi􀀂cation of AI in safetycritical systems, as the proposed framework provides a reliable and effective solution. Index Terms—aircraft braking, scienti􀀂c machine learning, certi􀀂cation, safety critical systems I. INTRODUCTION The current braking control system, reliant on rule-based logic, is nearing its operational limits due to the escalating complexity of aircraft systems over the last two decades. Furthermore, with the advent of next-generation aircraft, adhering to the traditional route of rule-based logic for executing such critical functions will become infeasible. Therefore, there is a pressing need to integrate an AI framework into the aircraft to undertake this task, which could be trained off-board in a safe manner. According to the statistical analysis of commercial aviation accidents by Airbus [1], over the last two decades, 12% of all fatal accidents and 59% of non-fatal hull losses occurred during 􀀃ight phases that rely heavily on landing gear and braking control performance. Reliable and resilient performance of the landing gear system in challenging environmental conditions, as well as in the face of system/component failures or malfunctions, is crucial for overall safety and the commercial feasibility of operations. However, current landing gear designs and braking control systems are limited to classical control strategies with little to no parametric adaptation andrecon􀀂guration capability, following a ”one solution 􀀂ts all”approach. It is also important to note that current designs are further disadvantaged by their minimal access to and usage of available aircraft sensing and health monitoring data. Aircraft braking control is a topic that has been discussed previously, yet there are several studies in the literature. The previous works have usually focused on designing an antiskid braking system. The 􀀂rst approaches were usually rulebased algorithms [2], and then the trend shifted to designing conventional controllers. Jiao et al. [3] designed a traditional controller for controlling the brake pressure, supported by runway identi􀀂cation and a neural network-based aerodynamic model. The same authors later proposed a heuristic controller to maximize the friction coef􀀂cient [4] . Another work done by D’Avico et al. observed aircraft dynamics through experimental 􀀃ight data and proposed a state machine-based control framework [5], [6] . Another study by Li [7] applied a standard PID controller for anti-skid. With the advancement of machine learning, some researchers used reinforcement learning to this problem. Liu et al. [8] proposed a twin delayed deep deterministic policy gradient algorithm, considering the aircraft’s 􀀂nal speed, simulation time, and slip ratio as reward functions. In a similar method, Radac [9] used standard Q-learning to tackle the problem. The braking control

FAQs

How does the proposed framework use scientific machine learning for landing braking?

It trains an agent to choose effective braking strategies under varying landing conditions while grounding the algorithm in landing-physics principles to ensure physically consistent outputs.

What performance does the framework achieve across runway conditions and landing speeds?

The results indicate the aircraft can decelerate successfully without skidding across all tested runway conditions and landing speeds.

How does the algorithm handle degraded brakes and initial yaw angles?

It maintains performance and safety even when brake performance degradation and initial yaw angles are introduced, improving robustness for safety-critical use.

Aug 5, 2026

Upload by Asher