Safety-aware predictive motion planning for close-range human-UAV collaboration in construction

Drones are increasingly used in construction for inspection and material transport, but their deployment in close-range collaboration with workers remains limited due to safety concerns and the difficulty of motion planning in dynamic environments. This paper introduces a predictive, risk-aware control framework integrating motion forecasting, probabilistic risk modeling, and hybrid planning to enable safe, efficient drone–worker interaction. Worker motion is captured with RGB-D input and forecasted 1.5 s ahead using PoseCastNet, a transformer-based network that outputs joint-wise 3D trajectories and confidence. Predictions are fused into a Bayesian-updated probabilistic safety map that informs global grid-based pathfinding and local actor-critic control with risk-sensitive rewards. Evaluations in simulation with occlusion and human motion yield a 96.5% success rate, over 40% improvement in minimum clearance, over 20% boost in task efficiency, and 8% reduction in joint prediction error compared to reactive and partially predictive baselines, demonstrating its effectiveness in enabling proactive, collaborative UAV operations.