N25A-T023 TITLE: Robotic Manipulation Learning for Repair of Marine Diesel Engines

OUSD (R&E) CRITICAL TECHNOLOGY AREA(S): Advanced Computing and Software;Human-Machine Interfaces;Trusted AI and Autonomy

OBJECTIVE: Develop autonomous robotic manipulation capabilities using artificial intelligence (AI) to repair and maintain marine diesel engines operating during long missions without human intervention. The overall objective of this topic is to provide the means to extend the duration of missions with uncrewed surface vessels and to provide greater assurance of mission completion.

DESCRIPTION: The Navy is developing a family of uncrewed surface vessels. Current designs utilize marine diesel engines for propulsion. These vessels are expected to conduct autonomous operations on long distance missions without sailors being placed aboard to maintain equipment. It would be desirable to have robotic systems onboard capable of inspection and basic maintenance tasks to enable these vessels to achieve the mission objective area. Due to bandwidth limitations, the ability to conduct teleoperated robotic manipulation is unlikely. There have been recent advances in AI and machine learning (ML) for robotic manipulation that exploit copious internet online video manipulation examples or learning from human demonstration, combined with foundation models and reinforcement learning to achieve autonomous manipulation. However, the manipulation tasks selected are often pick and place, or typical kitchen tasks. There is a need to transfer ML for manipulation to the more demanding tasks involved in equipment maintenance, which often involve tool use, significant torques, deformable (soft/squishy) object handling, difficult to access locations on equipment, or bimanual manipulations. This STTR topic focuses on manipulation, not on robotic locomotion. Maintaining marine diesel engines to sustain fully operational status involves a number of manipulation and inspection tasks. These manipulations include valve opening, rotation or depression of switches, rotation, removal or replacement of lids, adding fluids, adding or removing rubber gaskets and washers, catching liquids in containers, and tightening/untightening bolts, and reattaching loose components. Recent advances in computer vision have been made for object identification that will be needed for both manipulation and inspection. Inspection tasks include checking for deformations, cracks, leaks and loose components (brackets, bolts, hangers), recognition of warning lights, and reading gauges or status message text. This topic seeks the development of robotic manipulation technology capable of such tasks. This would include machine vision, robotic arms, and end effectors, with possible tool attachments. Operation at sea presents mechanical stability challenges, so design concepts are needed that would allow the robotic assembly to be stabilized, yet reach the essential engine components from the appropriate perspective. This could include sliding on a frame around the engine, or locomotion on a stable base.

PHASE I: Develop a design for robotic manipulation capable of the basic tasks involved in diesel engine maintenance, exploiting emerging research on ML of dexterous manipulation. This ML can be based on databases of video for engine repair manipulation tasks or on learning from human demonstration with reinforcement learning. The robotic manipulation concept includes designs or identification of robotic arms, end effectors, stable base and sensors for machine vision. Design a frame or base and means of shifting the position of the arm(s) to reach different areas of a diesel engine. Assess the feasibility of tool use, or end effector adaptations to accomplish basic elements of engine maintenance. Identify suitable cameras and machine vision processors capable of inspections that support diesel engine maintenance. Assess the feasibility of tactile or torque sensing to enable closed loop control of manipulators. The overall approach chosen should be mindful of the ultimate goal of autonomous manipulation, without teleoperation.

PHASE II: Develop prototype hardware and software to validate the concepts developed in Phase I. This can include commercial off-the-shelf (COTS) robotic components. Demonstrate the learning and execution of maintenance relevant robotic manipulation tasks on a diesel engine, or close surrogate. Possible tasks include valve opening, rotation or depression of switches, rotation, removal or replacement of lids, adding fluids, adding or removing rubber gaskets and washers, catching liquids in containers, and tightening/untightening bolts, and reattaching loose components. While the training of this system can involve teleoperation, the final demonstration should show autonomous manipulation to execute the sample tasks. Develop a prototype machine vision system to work with the robotic manipulators for basic inspection tasks and for closed loop control of the manipulation task.

PHASE III DUAL USE APPLICATIONS: Demonstrate and evaluate the robotic manipulation technology at a Naval LBES facility with marine diesel engines and on board a Navy uncrewed surface vessel with a marine diesel powerplant in coordination with PMS 406.

This technology has strong commercial potential since many commercial vessels use marine diesel engines. This technology would also be applicable to fleets of trucks using diesel engines.

REFERENCES:

1. Spencer, J.; Choudhury, S.; Barnes, M.; Schmittle, M.; Chiang, M.; Ramadge, P. and Srinivasa, S.S. "Expert intervention

learning: An online framework for robot learning from explicit and implicit human feedback." Autonomous Robots, Volume 46, Issue 1, 2022, pp. 99-113. https://dl.acm.org/doi/10.1007/s10514-021-10006-9

2. Levine, S.; Kumar, A.; Tucker, G. and Fu, J. "Offline reinforcement learning: Tutorial, review, and perspectives on open problems." arXiv preprint. arXiv:2005.01643, 2020.2021202220232024

3. Deshpande, Abhay; Ke, Liyiming; Pfeifer, Quinn; Gupta, Abhishek and Srinivasa, Siddhartha S. "Data Efficient Behavior Cloning for Fine Manipulation via Continuity-based Corrective Labels." IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024. https://arxiv.org/abs/2405.19307v2

4. Zhao, Tony Z.; Kumar, Vikash; Levine, Sergey and Finn, Chelsea. "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware." https://doi.org/10.48550/arXiv.2304.13705

5. Bahl, Shikhar; Mendonca, Russell; Chen, Lili; Jain, Unnat and Pathak, Deepak. "Affordances from Human Videos as a Versatile Representation for Robotics." CVPR 2023. https://arxiv.org/abs/2304.08488

6. Arunachalam, Sridhar Pandian; Silwal, Sneha; Evans, Ben and Pinto, Lerrel. "Dexterous Imitation Made Easy: A Learning-Based Framework for Efficient Dexterous Manipulation." ICRA 20. https://arxiv.org/abs/2203.13251

KEYWORDS: Robotics; manipulation; maintenance; machine learning; diesel engines; machine vision

TPOC 1: Thomas McKenna
Email: [email protected]

TPOC 2: Bob Brizzolara
Email: [email protected]

---