A predictive risk-aware dynamic window approach for USV obstacle avoidance in dynamic environments

Unmanned surface vehicles (USVs) operating in dynamic-obstacle environments often suffer delayed avoidance and fluctuating controls because future encounter risks are insufficiently anticipated. This paper proposes a predictive risk-aware dynamic window approach (PR-DWA) to improve proactive risk representation and online local planning. PR-DWA retains the velocity-space sampling and receding-horizon optimization framework of the classical dynamic window approach (DWA), while incorporating short-term obstacle-state prediction, uncertainty-inflated safety screening with a conservative Gaussian probability bound, and conditional value at risk (CVaR)-based tail-risk evaluation. Future obstacle states and uncertainty are represented in mean-covariance form. Candidate trajectories are first screened by the probability-bounded condition, and feasible trajectories are ranked using tail risk together with goal convergence, navigation efficiency, and control smoothness. Numerical simulations demonstrate that PR-DWA achieves the highest composite score in all six scenarios under fixed-reference normalization and maintains positive minimum clearance in each case. Its total near-risk exposure time over the six scenarios is 5.15 s, compared with 14.05 s for classical DWA. The conservative screening condition increases the mean path length by 0.123% relative to Monte-Carlo probability screening, while reducing the average computation time from 19.583 ms to 2.266 ms. A matched stochastic extension and component ablation further evaluate robustness and the contribution of uncertainty screening. Hardware-in-the-loop simulation and field experiments confirm online avoidance with real observations.