Disturbance-Adaptive Differentiable MPC for Underwater Structure Inspection Using Underwater Robot
Minjong Kim, Bonchul Ku, Dongsub Kim, Young-woon Song, WOOJIN SEOL, You hyun Jang, Son-Cheol Yu
AI summary
Problem
Fixed-weight MPC cannot adapt to non-stationary underwater disturbances, while model-free RL lacks safety guarantees and stability, hindering reliable close-proximity inspection of marine infrastructure.
Approach
The framework embeds a differentiable iLQR-based MPC layer inside an actor-critic network that learns time-varying cost weights online to adapt to flow and payload disturbances.
Key results
- 100% orbit success rate under randomized disturbances
- Zero safety constraint violations across all horizons
- Superior trajectory smoothness and reliability over fixed MPC and PPO
- Robust performance maintained even with short prediction horizons
Why it matters
Enables safe, reliable autonomous inspection of underwater infrastructure in dynamic, real-world conditions where traditional controllers fail.
Abstract
Close-proximity inspection of underwater cylin- drical structures is challenging due to nonlinear vehicle dy- namics, flow disturbances, and payload uncertainty. Fixed- weight MPC provides structured constraint handling but lacks adaptivity, while model-free RL is adaptive but often unstable and unsafe under disturbances. We propose Marine AC-MPC, which combines a differentiable iLQR-based MPC layer with an actor-critic framework that learns time-varying MPC cost weights online. In MarineGym, the proposed method achieves more reliable orbit tracking and higher success rates than fixed- weight MPC and PPO baselines under disturbed conditions.