Learning Scalable Salp-Inspired Locomotion

Manuel Agraz Vallejo, Kagan Tumer

Keywords: Reinforcement Learning, Continuous Control, Graph Neural Networks

TL;DR: In this work we introduce a novel learning environment and train a set of graph-based controllers for salp-inspired agents (simple thrust-based agents that can form chains for locomotion).

Abstract: Salp-inspired robots have the potential to monitor marine life and collect scientific samples in topologically intricate underwater habitats, such as caves, overhangs, and confined openings. Their chain-like structure enables them to detach and reattach salp units, allowing them to disassemble and navigate through narrow bottlenecks, and then reassemble to reach otherwise inaccessible areas. However, the redundancy and non-linear characteristics of multi-jet propulsion, along with the chain structure, make designing modular locomotion controllers challenging. Current approaches focus on controlling single salp units or small salp chains and are limited to operating on a fixed salp chain structure. In this work, we propose the use of Reinforcement Learning methods to learn a controller that can handle a variable input and output space, enabling locomotion across multiple salp chain lengths. We perform a comparative study among graph-based models to determine the most well-suited architecture for a novel Salp Chain Locomotion Domain. Our methodology yields a set of policies that enable salp chain locomotion across various chain lengths, and our analysis demonstrates that modeling the salp chain state as a chain graph structure yields comparable zero-shot performance to modeling it as a fully connected graph.

Journal Edition Interest: Yes

Submission Number: 35