@Article{iasc.2026.088941,
AUTHOR = {Bao Nguyen, Ngan Nguyen Xuan Phuong},
TITLE = {A Hybrid Diffusion World Model for UAV Trajectory Forecasting and Collision Risk Estimation in Dense 3D Environments},
JOURNAL = {Intelligent Automation \& Soft Computing},
VOLUME = {41},
YEAR = {2026},
NUMBER = {1},
PAGES = {105--137},
URL = {http://www.techscience.com/iasc/v41n1/68855},
ISSN = {2326-005X},
ABSTRACT = {Autonomous unmanned aerial vehicles (UAVs) operating in dense three-dimensional environments require predictive models that can represent multiple plausible futures while estimating the safety consequences of these futures. This paper presents a hybrid diffusion world model for short-horizon UAV trajectory forecasting and probabilistic collision-risk estimation. The model conditions on historical UAV states and executed actions, depth observations, and safety-context variables to generate multiple future relative-motion trajectories over a 1.0-s prediction horizon, while jointly estimating collision probability, near-miss probability, and obstacle-clearance information. A task-specific synthetic UAV dataset based on locations in Vietnam containing 1000 in-distribution episodes and 200,000 synchronized frames is generated under randomized static and dynamic obstacle configurations, controller behaviors, sensing noise, and disturbances. The dataset is partitioned exclusively at the episode level into 697 training, 148 validation, and 155 test episodes, and all baselines and ablations use the same immutable partition and preprocessing protocol. To improve statistical rigor, all model families are trained using five independent random seeds, and evaluation is performed on the complete test partition rather than on an ordered subset of sliding windows. Statistical uncertainty is quantified at the episode level, while probabilistic risk estimation is evaluated using Brier score, expected calibration error, and reliability diagrams with classification thresholds selected exclusively from the validation partition. The diffusion model is additionally evaluated using multiple sampled futures with <math id="mml-ieqn-1"><mi>K</mi><mo>=</mo><mo fence="false" stretchy="false">{</mo><mn>5</mn><mo>,</mo><mn>10</mn><mo>,</mo><mn>20</mn><mo fence="false" stretchy="false">}</mo></math>, including mean and best-of-<math id="mml-ieqn-2"><mi>K</mi></math> trajectory errors and trajectory-diversity measures. Matched recurrent baselines, deterministic prediction, risk-only prediction, and ablations removing depth, safety context, and risk heads are included to isolate the contribution of individual components. The results show that deterministic recurrent models remain competitive in single-trajectory ADE and FDE, whereas the diffusion model provides multimodal future coverage and substantially improved probabilistic calibration of collision risk. These findings support diffusion-based world modeling as a complementary predictive safety layer for uncertainty-aware UAV navigation rather than solely as a point-trajectory regressor.},
DOI = {10.32604/iasc.2026.088941}
}