Abstract:
To address the challenge of simultaneously ensuring tracking accuracy, safe bottom clearance, and compliance with actuator constraints for an autonomous undersea vehicle(AUV) operating near complex seabed terrain, this paper proposes a terrain-following method based on sliding-window B-spline planning and a meta-learning-based, weight-adaptive linear time-varying model predictive control(LTV-MPC) strategy. The planning layer employs cubic B-splines subject to slope constraints to generate a reference path that maintains a prescribed altitude margin, while a sliding-window mechanism balances local optimization with global continuity. The control layer implements an LTV-MPC controller equipped with a meta-learning network that adaptively adjusts the cost-function weights online based on features such as terrain curvature, depth error, and attitude error, thus accommodating variations in tracking difficulty across different terrain conditions. The meta-network parameters are trained offline through Bayesian optimization using closed-loop simulation data from multiple synthetic terrains, enabling the network to learn the mapping between terrain features and optimal weights. In the online stage, only forward inference is required, resulting in high computational efficiency. In the specified simulation scenarios, validation using synthetic and measured terrain data demonstrates that the proposed method reduces the root-mean-square(RMS) tracking error by 42.5% and 31.3%, respectively, compared with line-of-sight guidance combined with proportional-integral-derivative control(LOS-PID). Compared with standard LTV-MPC, the RMS tracking error is reduced by 4.5% and 6.1%, respectively. Meanwhile, the pitch-angle and rudder-angle constraints are satisfied, effectively resolving the trade-off between tracking accuracy and safety under complex terrain conditions.
