Restoring complex movement and locomotion after paralysis through collaborative copilots

The National Institutes of Health Office of the Director awarded the University of California, Los Angeles $2.205 million under the Trans-NIH Research Support program (CFDA 93.310) on September 30, 2025, for research on restoring complex movement and locomotion after paralysis through collaborative AI copilots.

The project develops low-risk, high-performance assistive devices for individuals with paralysis by training artificial intelligence copilots to infer user intent and enable performance of complex everyday movements such as folding laundry, preparing food, opening door knobs, and grasping objects. The research encompasses three technical thrusts: training copilot architectures that combine computer vision with non-invasive signals reflecting user intent to perform complex actions; fusing multiple non-invasive signal sources to provide sufficient information for the copilot to infer user goals and assist task completion; and developing adaptive methods to improve system performance and robustness as the user and copilot share control.

Performance occurs at UCLA's Los Angeles campus (90095) through August 31, 2030. The award is classified as a project grant under the Trans-NIH Research Support program, which supports research across multiple offices within the NIH Office of the Director and undergoes peer review for scientific merit and program relevance with a review cycle of four to nine months.

Abstract

There are no widespread devices available that significantly improve the quality of life for those with paralysis. A widespread and effective device for paralysis should be low risk, yet high performance, enabling complex movements needed to perform everyday tasks. We propose to achieve a low risk, high performance device through three thrusts. Our first thrust is to capitalize on advances in artificial intelligence (AI) that enable US to train robots that perform the same types of tasks that humans perform, including folding laundry, putting food in a pan, opening door knobs, or picking up a bag of chips. Because robots can perform these tasks, we reason that as long as an AI copilot can infer a human's goal to perform a certain movement, the copilot can then help perform the complex movements associated with the task. We achieve this by training new copilot architectures that use computer vision, as well as non-invasive signals reflecting the user's intent, to perform complex actions. Our second thrust is to take advantage of and fuse many non-invasive signal sources, which together provide enough information for the copilot to infer the user's goal and help the user complete an action. Our third thrust is to develop methods to adapt the user and copilot as they share control. We hypothesize this will have a key impact on increasing system performance and robustness. We believe this proposed work, if successful, will have a transformative impact on the development of clinically viable devices that help people with paralysis move autonomously.