Video Recording of the MS Thesis Defense on "Safe and Agile UAV Navigation in Confined and Tight Spaces via Deep Reinforcement Learning" by Mr. Ziad Shoeib, a Fulltime Graduate Student of CIE Dept., KFUPM, Dhahran, Saudi Arabia, May 6th, 2026.

Abstract:

Small UAVs are increasingly deployed for inspection in confined infrastructure such as ventilation ducts and underground service tunnels, where agile flight must be reconciled with tight clearances and strong proximity-induced aerodynamic disturbances. This thesis develops a learning-based, safety-aware framework for agile quadrotor navigation in narrow circular tunnels, trained with Proximal Policy Optimization (PPO) in a custom PyBullet environment. At its core, the framework couples a smoothed log-barrier centerline penalty with an Estimated Disturbance Level (EDL)-modulated barrier in which a Computational Fluid Dynamics (CFD)-trained Multilayer Perceptron (MLP) dynamically tightens the safe corridor in aerodynamically demanding sections, unifying geometric and aerodynamic safety into a single signal. Training-time Exponential Moving Average (EMA) filtering of the policy’s actions proves essential for suppressing bang-bang motor commands; the resulting flagship configuration completes 95/95 waypoints at 2.23ms−1 with 4.2cm mean centerline deviation, reducing simultaneous motor saturation from 84.5% to 0.7%.

Speaker Bio:

Ziad Shoeib earned his M.Sc. in Systems and Control Engineering from King Fahd University of Petroleum and Minerals (KFUPM), Saudi Arabia, in 2026, after receiving his B.Sc. in Mechatronics Engineering from Misr University for Science and Technology (MUST), Egypt, in 2023. His research interests lie at the intersection of aerial robotics, control theory, and machine learning, with a focus on safe reinforcement learning for autonomous robotic systems. His master’s thesis developed a PPO-based, safety-aware framework for agile quadrotor navigation in narrow tunnels, integrating log-barrier safety shaping, aerodynamic disturbance awareness, and EMA-based action smoothing to support stable flight in tight-clearance conditions. His background also includes experience in robotic simulation, experimental UAV systems, and teaching in robotics and control-related courses.