Learning Manipulation Skills for Marine Diesel Engine Repair

The Navy's modernization efforts prioritize the development of uncrewed surface vessels (USV) capable of autonomous long-distance missions. A central challenge facing this development is performing the required maintenance and repair of marine diesel engines, which power envisioned USV, operating without human intervention. While the Navy plans to implement robotic systems to perform maintenance and repair needs, bandwidth and latency limitations in operational areas are expected to preclude solutions involving remote teleoperation. Innovations in artificial intelligence and machine learning have shown potential for automating manipulation tasks. However, existing approaches require significant data and are challenged by the complexity of tasks involved in engine maintenance. To address this challenge, we propose a solution that focuses on developing atomic, reusable manipulation skills that can be sequenced into complex tasks, leveraging the Continuity based Corrective Labels for Imitation Learning (CCIL) and Affordance Template frameworks. We call the resulting system CLOVER (Composing Learned Operations for Various Engine Repairs). The resulting system unifies a novel learning framework with an autonomy approach vetted in commercial applications to mitigate data scarcity issues and the impact of covariate shift.