Inertial Measurement Unit (IMU)-based Pedestrian Dead Reckoning (PDR) enables infrastructure-free indoor positioning and requires heading-drift compensation associated with gyroscope bias. This paper proposes a stationary-window-based Deep Belief Network (DBN) framework that learns a gyroscope bias representation from z-axis angular-velocity and sampling-interval sequences observed during stationary intervals and applies it to heading correction during walking. The learned representation captures the residual angular-velocity offset under stationary conditions and serves as an adaptive correction term for subsequent heading integration. Experiments on short-term, three-lap long-term, and complex indoor paths show that stationary-window-based DBN bias estimation is particularly effective in accumulated-drift regimes, such as long-term repeated walking and complex multi-turn trajectories. The proposed DBN method achieved an average absolute trajectory error (ATE) of 3.9587 m in the long-term experiment and 0.9126 m in the complex path experiment. Stationary detection sensitivity analysis further shows that false-positive stationary decisions have a stronger influence on trajectory consistency than false-negative stationary decisions. These results show that the proposed framework maintains stable waypoint-level trajectory behavior in long-term and complex PDR scenarios where heading drift becomes more pronounced.

Conventional Low-Rank Adaptation (LoRA) constrains weight updates to a static linear low-rank manifold, which is inherently limited when applied to Reinforcement Learning (RL) tasks for Unmanned Aerial Vehicle (UAV) applications. UAVs operate in highly dynamic and nonstationary environments where rapid variations in sensing and state transitions lead to complex, nonlinear input–output relationships. Such environmental complexity cannot be adequately modeled by a static Low-rank approximation, making conventional LoRA approaches insufficient for the high-dimensional dynamics required in UAV applications. To overcome these limitations, we propose an attention-enhanced LoRA that constructs an input-dependent and intrinsically nonlinear adaptation manifold. By integrating a nonstandard attention mechanism into the vanilla LoRA, our method enables the model to dynamically reshape its weight subspace in response to changing environmental conditions. This allows the policy and value networks to capture diverse local patterns as well as global contextual structure during adaptation, ultimately improving robustness under domain shift and nonstationary data distributions. We evaluate the proposed method in UAV adaptation scenario based on the AirSim simulator, where a multi-agent training is conducted with internally collected datasets, including multi-sensor observations and UAV physical state information, and policies are transferred from obstacle-free to cluttered environments. Compared to vanilla LoRA, the proposed method reduces initial reward variance by over 70%, leading to earlier adaptation and more stable generalization, and exhibits richer nonlinear expressive power, allowing the model to accommodate the complex, high-dimensional characteristic of UAV tasks.