Harvard seminar explores how AI learns complex physical movements

Researchers are using musculoskeletal simulations and machine learning to build AI agents capable of reproducing complex physical movements, according to a Kempner Institute seminar. EPFL professor Alexander Mathis presented a decade of progress in the field, including its connections to human motor learning.
- Alexander Mathis presented a decade of musculoskeletal AI research at a Kempner seminar.
- Annual competitions since 2017 have tested AI controllers on simulated physical tasks.
- Researchers use imitation learning and reinforcement learning to train the agents.
- Performance dips may reflect stages of human motor learning.
- Potential applications include prosthetics, rehabilitation and assistive technologies.
Training AI to control muscles and joints
Mathis, an associate professor at the Brain Mind Institute at EPFL, discussed the challenge of creating AI that can reproduce movements such as returning a table-tennis serve or scoring a soccer goal. These tasks require precise coordination across multiple joints and muscles, making them difficult for conventional AI systems to perform.
Over the past decade, researchers have used increasingly sophisticated musculoskeletal simulations to study how AI agents navigate and interact with dynamic physical environments. Mathis presented this work at a recent Kempner Seminar Series talk titled The Neuroscience of Skill: Lessons from a Decade of Musculoskeletal AI.
Competitions test simulated physical skills
Mathis traced the development of these agents through annual competitions held each year since 2017 in conjunction with the Conference on Neural Information Processing Systems. The competitions ask researchers to create AI controllers for physiologically realistic musculoskeletal models.
The simulated challenges include scoring a penalty kick and manipulating Baoding balls. Success depends on coordinating several parts of a simulated body at once, rather than simply selecting an answer or recognizing an image.
Researchers have used two complementary machine-learning methods to train the agents. In imitation learning, an agent learns by copying a demonstration of a successful movement. In reinforcement learning, it pursues a goal through trial and error.
Performance dips may resemble human learning
As the models improved, their performance sometimes dipped before rising to a higher peak. Mathis links this pattern to a theory of motor learning proposed in 1967 by American psychologists Paul Fitts and Michael Posner.
The theory describes three stages of skill development. During the cognitive stage, a learner consciously considers the steps required for a task. In the associative stage, that knowledge is translated into how to perform the movement. In the autonomous stage, the skill can be applied with minimal conscious control.
According to Mathis, the temporary declines in AI performance may resemble the way human learners improve their technique. By breaking movements into conscious steps, learners may temporarily return to the cognitive stage before developing a stronger performance.
The research could support the development of more dexterous and agile robots while also offering insight into how humans acquire motor skills. The Kempner Institute says potential applications include prosthetics, rehabilitation and assistive technologies.
Kempner Institute (Harvard): AI on the Move


