Eager: toward end-to-end control for robotic lower-limb prostheses with a validated simulation approach

The National Science Foundation Division of Computer and Network Systems awarded the Regents of the University of Michigan $356,515 on October 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop end-to-end control strategies for robotic lower-limb prostheses using simulation and artificial intelligence.

The project will validate a high-fidelity biomechanical simulation model of a user-prosthesis system and develop a foundational dataset of people with amputations performing cyclic and non-cyclic tasks with prostheses, collected across multiple laboratories. This dataset will serve as training data for sim-to-real data-transfer networks and enable development of an end-to-end control framework that combines supervised learning with off-policy reinforcement learning to close the sim-to-real gap in sensor data and estimate joint torques for prosthetic control. The work will demonstrate the framework's effectiveness using the Open-Source Leg, an open-source platform for shared control development.

Work is performed at Ann Arbor, Michigan. The period of performance runs from October 1, 2026, through September 30, 2027. This is a project grant, classified as an EAGER award to support exploratory and preliminary research.

Abstract

The long-term objective of this project is to leverage the simulation and artificial intelligence (AI)-enabled workflow of modern robotics, enabling the generalization of these methods to the unique and data-scarce application of robotic prostheses. Developing generalizable learning-based control strategies for robotic prostheses will require both high-fidelity simulation models and large-scale datasets, resources that remain scarce, difficult, and expensive to obtain. Despite significant advances in simulation that have enabled the development of some large human locomotion datasets, there remains a critical need to characterize and reduce discrepancies between simulated and physical system behavior for prosthesis users. We will demonstrate the framework's effectiveness by applying IT to the open-source leg, an open-source platform that enables fast, cross-cutting advances in shared control development. This initiative will provide critical training data for advanced simulation and end-to-end deep learning and reinforcement learning control networks to function and will provide initial feasibility of such approaches. The work will validate a high-fidelity simulated biomechanical model of a user-prosthesis system that can be leveraged for end-to-end control development. This work will build a foundational dataset of people with amputations performing cyclic and non-cyclic tasks with the prosthesis collected across laboratories. The dataset will serve as the baseline for training the SIM-to-real data-transfer network. Additionally, this work will validate development of an end-to-end control framework that closes the SIM-to-real GAP in sensor data and estimates joint torques for prosthetic control, combining supervised learning with off-policy reinforcement learning. The outcome of this project will include a validated foundation to leverage physics-based simulation in the development of a new generation of prosthesis control approaches at the forefront of robotics. 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.- subawards are planned for this award.