Layout-independent actuation allocator for fin-actuated marine robots

2026-07-03

AI Summary

This work addresses the challenge of redesigning control allocators for fin-driven marine robots whenever actuator layouts change. To overcome this limitation, the authors propose a layout-agnostic, zero-shot cross-layout control method that leverages graph neural networks and Transformers to model the robot’s geometric structure. A mixture density network predicts a multimodal distribution over control commands, which are subsequently refined at inference time using a differentiable physics-informed surrogate model to minimize trajectory tracking error and energy consumption. The approach achieves, for the first time, zero-shot generalization across diverse out-of-distribution actuator configurations using a single unified model. Experimental validation in a real-world pool environment demonstrates tracking accuracy comparable to that of layout-specific controllers.

Abstract

In this study, we propose a layout-independent control allocator capable of zero-shot deployment across diverse actuator configurations. The proposed method utilizes a learning pipeline that integrates a Graph Neural Network (GNN) and a Transformer to represent the robot's geometric layout as a graph, along with a Mixture Density Network (MDN) to predict multi-modal control command distributions. Furthermore, by incorporating a differentiable physics surrogate model, we achieve command refinement during inference to minimize target wrench tracking error and energy consumption. A single generalized model using randomly generated actuator layout data demonstrated high trajectory tracking performance on different actuator layout robots outside the training distribution. Additionally, in real-world pool experiments, our approach achieved performance nearly equivalent to conventional controllers designed to specific layouts.

Problem

• layout-independent
• control allocation
• fin-actuated marine robots
• zero-shot deployment
• actuator configuration

Innovation

• layout-independent control
• Graph Neural Network
• Mixture Density Network
• differentiable physics surrogate
• zero-shot deployment