Stability-constrained incremental learning for robust ROV control with feasibility-guided data selection

Remotely operated vehicles (ROVs) often operate in complex aquatic environments with uncertain disturbances, while acquiring high-quality demonstrations is expensive and limited in practice. This makes it challenging to improve control policy in a stable and data-efficient manner. To address this issue, this paper proposes a stability-constrained incremental learning framework for high-level ROV control. A baseline policy is first constructed via an incremental learning scheme under a Lyapunov-inspired structural constraint, which enforces a stability-preserving policy parameterization throughout the model growth process. This allows the policy to be progressively improved without violating the underlying stability structure, even when only limited demonstrations are available. To further enhance learning from sparse data, a behavioral feasibility predictor is developed to support data-efficient policy refinement under disturbances. Instead of enforcing hard safety constraints, the predictor reshapes the effective training distribution by filtering out behaviorally inconsistent data, enabling the policy to reuse historical interaction data and improve robustness under unmodeled disturbances. A stability analysis is provided for the constrained policy structure, and a monotonic approximation property is established for the incremental learning procedure. The proposed framework is evaluated in simulation and validated on a real ROV platform in planar trajectory tracking tasks under external disturbances. Experimental results demonstrate improved tracking robustness and behavioral consistency compared with conventional learning-based control approaches under limited demonstrations.