Transformer-Guided Deep Reinforcement Learning for Trajectory Design
Advisor: Xiaosong Du, xiaosongdu@mst.edu
Description
Electric vertical takeoff and landing (eVTOL) aircraft are limited in range by battery capacity, making energy-efficient takeoff trajectory design critical for practical operation. While deep reinforcement learning (DRL) enables adaptive control, it often struggles to discover feasible solutions due to highly nonlinear dynamics, strict constraints, and high sample complexity. This work introduces a transformer-guided deep reinforcement learning (TDRL) framework for eVTOL takeoff trajectory design. A transformer trained on optimal control trajectories learns temporal relationships in control sequences and guides policy exploration by restricting actions to feasible, energy-conscious regions. The approach is evaluated across a design space varying aircraft parameters, namely efficiency and wing planform scale. Transformer-guided agents achieved feasible trajectories at all evaluated design points, maintaining energy consumption within 5% of optimal solutions, whereas vanilla DRL failed in most cases. Results demonstrate that transformer guidance improves training reliability and performance across vehicle configurations.
Transformer-Guided Deep Reinforcement Learning for Trajectory Design
Havener Center, Miner Lounge / Wiese Atrium, 1:30pm-3:30pm
Electric vertical takeoff and landing (eVTOL) aircraft are limited in range by battery capacity, making energy-efficient takeoff trajectory design critical for practical operation. While deep reinforcement learning (DRL) enables adaptive control, it often struggles to discover feasible solutions due to highly nonlinear dynamics, strict constraints, and high sample complexity. This work introduces a transformer-guided deep reinforcement learning (TDRL) framework for eVTOL takeoff trajectory design. A transformer trained on optimal control trajectories learns temporal relationships in control sequences and guides policy exploration by restricting actions to feasible, energy-conscious regions. The approach is evaluated across a design space varying aircraft parameters, namely efficiency and wing planform scale. Transformer-guided agents achieved feasible trajectories at all evaluated design points, maintaining energy consumption within 5% of optimal solutions, whereas vanilla DRL failed in most cases. Results demonstrate that transformer guidance improves training reliability and performance across vehicle configurations.
