Incorporating Pitching Moment Observation in Reinforcement Learning for Enhanced Thermal Updraft Exploitation by Fixed-Wing UAVs

Fixed-wing UAVs can extend endurance by exploiting atmospheric thermals. Prior reinforcement learning (RL) methods used vertical acceleration and roll moments as state inputs. However, when flying straight through a thermal core, roll cues are weak, making it hard for RL agents to anticipate circling. This can delay thermal centering and waste potential energy gain. We propose adding pitching moment as an additional observation. Pitching moment reflects vertical airflow gradients along the fuselage, providing early cues of strong up-drafts. We design a deep RL agent with three-layer neural network and discrete bank angle actions. Simulations use a stationary Gaussian thermal model and glider dynamics. Results show that the pitchaugmented agent recognizes thermals earlier and begins circling sooner. Compared to baseline models, it gains up to ∼ 8% more altitude in head-on approaches and achieves better centering stability. This confirms that pitching moment sensing enhances autonomous soaring. Future work will explore moving thermals and real-world flight tests.