IROS26
Layout-independent actuation allocator for marine robots
Yuya Hamamatsu, Maarja Kruusmaa, Asko Ristolainen
Tallinn University of Technology
Code (Coming soon) arXivAbstract
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.
Video
Overview
This study introduces a layout-independent control allocator for underwater robots that enables zero-shot deployment across various actuator configurations without requiring platform-specific redesigns.
The approach represents the robot’s fin placement and orientation as a graph using a Graph Neural Network (GNN) and a Transformer. A Mixture Density Network (MDN) then predicts optimal control command distributions. Additionally, a differentiable physics surrogate model refines these commands during inference, minimizing both trajectory tracking errors and energy consumption.
Evaluated through both simulations and real-world pool experiments using a single model trained on randomly generated layouts, the system achieved tracking performance comparable to traditional layout-specific controllers. It also demonstrated robust fault tolerance in scenarios where one or two fins failed, and successfully reduced mean power consumption by 11.9% compared to baseline methods.