CRII: RI: Towards Robust and Efficient Bipedal Robot Locomotion on the Moon through Reinforcement Learning

Michigan Technological University, Houghton MI

Investigators

Abstract

Beyond Earth, planetary exploration is still heavily reliant on wheeled rovers, which show limited mobility when it comes to difficult terrain and inclined regions, for instance, featuring rocks, craters, narrow crevices. Bipedal robots hold promise for expanding the scope of tasks on the Moon due to their unique locomotion capabilities. Because of their human-like structure, a bipedal robot can also mimic some of the movements and actions of astronauts, allowing for better testing of equipment and procedures for humans. However, the challenges of lunar gravity and the presence of fine surface’s dust pose stability challenges for these robots. This award seeks to fund research that attempts to gain a fundamental understanding of lunar bipedal locomotion mechanics and design robust and energy-efficient devices for bipedal robot locomotion on the Moon by using reinforcement learning. This work has significant potential to impact applications on challenging terrain on Earth and has a translational impact on improving human locomotion strategies on the Moon. This will lead, in turn, to better mission planning and increased astronaut safety. This research project will contribute to build and improve the robotics curriculum at Michigan Technological University while engaging diverse groups, including low-income, first-generation, and/or underrepresented undergraduate and graduate students, high school students, in robotics education and research. The research goal of the project is to leverage dynamic modeling, control theory, hybrid systems, and reinforcement learning to advance biped robot locomotion on the Moon. To achieve this goal, three research objectives will be pursued: i) improving the understanding of lunar biped locomotion mechanics by using both abstraction and detailed models; ii) investigating robust and energetically efficient biped locomotion control on the Moon by using high-fidelity simulation and reinforcement learning; and iii) designing and constructing a lunar simulation platform and evaluate the performance of the identified controllers through experiments. The fundamental understanding of lunar locomotion mechanics can provide valuable insights into the interactions between robots and lunar terrain and inform future missions and robot design. Control theory can help explain the locomotion mechanics and accelerate training efficiency while machine learning enables the discovery of novel locomotion strategies on the Moon. Additionally, the high-fidelity simulation and the reinforcement algorithm have potential applicability in diverse terrains, including compliant and soft surfaces, thereby advancing the mobility capabilities of state-of-the-art bipedal robots. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.