In-Context Meta-Learning-Based Model Predictive Contouring Control for Unmanned Surface Vehicles

This article proposes an in-context meta-learning-based model predictive contouring control (INCML-MPCC) framework for unmanned surface vehicles (USVs). To address the distinct dynamics across different operating speeds, we decouple the USV dynamics model into two distinct components: a shared neural network backbone and a low-dimensional context vector. Specifically, the backbone is optimized to capture shared hydrodynamic structures across diverse speed intervals, while the context vector is designed to explicitly encode task-specific residual variations. In the offline phase, restricting inner-loop adaptation to the context vector acts as an intrinsic regularizer, preventing overfitting to specific tasks. In the online phase, this adaptive model is integrated into the MPCC optimization as a predictive model. Here, the backbone remains frozen to preserve general physical knowledge while the context vector is updated via a replay buffer to dynamically capture local hydrodynamic residuals. This strategy drastically reduces the number of trainable parameters and mitigates the risk of catastrophic forgetting, ensuring robust generalization to unseen operating conditions. Simulation and hardware experiments demonstrate that INCML-MPCC achieves superior tracking performance compared with the existing benchmarks.