On May 23, Mark Zuckerberg announced that researchers from the Meta Artificial Intelligence Research Laboratory and the Neuromechanical Modeling and Engineering Laboratory at the University of Twente in the Netherlands, led by Professor Massimo Sartori, had collaborated on the development from an open source framework called “MyoSuite”. which combines advanced musculoskeletal models with advanced AI. MyoSuite’s AI-driven digital models can learn complex movements and interactions with assistive robots, which would require extensive experimentation on real human subjects. They titled their study “ Myosuite, a contact-rich simulation suite for musculoskeletal motor control ».
The motor behavior of an intelligent being is complex. To apprehend him, doctors Vikash Kumar and Vittorio Caggiano of Meta AI Research joined forces with the
Professor Massimo Sartori and Drs Huawei Huang and Guillaume Durandau from the University of Twente, Enschede, The Netherlands, to develop MyoSuite: a set of musculoskeletal models and tasks that enable the application of ML to solve biomechanical control problems. MyoSuite combines the two faces of intelligence: motor and neural.
Biomechanics aims to study the functions and properties of movement of the human body, it is very useful for creating equipment adapted to the human body (seats, desks, cars, etc.), improving sports performance, in the field of health etc .
However, human biomechanics is highly complex and requires efficient coordination between the central nervous system, where decisions are synthesized by networks of billions of neurons, and the peripheral musculoskeletal system, which translates these intentions into actions.
Exploring musculoskeletal motor control through ML
ML algorithms are rarely used to explore complex motor control situations, such as musculoskeletal control.
There are in silico frameworks that, like OpenSim, contain physiologically detailed musculoskeletal models, however they do not have the capacity for complex interaction with the physical world outside the agent’s body. According to the researchers, “These existing frameworks are not built on complex and skillful motor tasks nor are they computationally efficient or scalable enough to meet the data needs of ML algorithms. MyoSuite fills these gaps. »
The MyoSuite platform was designed to study the physiological details behind musculoskeletal motor control. It includes a comprehensive set of accurate musculoskeletal models that take into account musculoskeletal dynamics and their temporal interactions, such as muscle fatigue or sarcopenia, which affects humans around the age of 50, when muscle mass and strength decline significantly, leading to, for example, gait problems, as well as behavioral tasks of daily living, injury rehabilitation, and prosthetic/exoskeleton assistance.
According to the team, MyoSuite’s musculoskeletal models are up to 4,000 times faster than other simulators to meet the data demands of modern ML algorithms.
Professor Massimo Sartori says:
” All of this is achieved by combining state-of-the-art musculoskeletal models with state-of-the-art artificial intelligence for movement behavior synthesis.. »
real world applications
MyoSuite synthesizes behaviors, but can also facilitate applications with real-world implications, such as rehabilitation, surgery, and shared autonomy assistive devices.
The researchers took the example of a tendon tear, MyoSuite modeled tendon transfer, a common technique for regaining functionality due to a torn tendon. He then simulated the outcome of the surgery and the impact it will have on functional rehabilitation.
This research could have a significant impact for the development of prosthetics and post-traumatic rehabilitation, because MyoSuite generated physiologically realistic movements, such as turning a pen or manipulating Baoding balls, in great detail. The team has released it as open source and will soon launch MyoChallenge, a NeurIPS competition track where the ML community will be invited to participate in solving two of the most difficult dexterity challenges: reorienting the die and simultaneously spinning the two balls. from Baoding.
MYOSUITE: A CONTACT-RICH SIMULATION SUITE FOR MUSCULOSKELETAL MOTOR CONTROL
Vittorio Caggiano (Meta AI Research), Huawei Wang (University of Twente), Guillaume Durandau (University of Twente), Massimo Sartori (University of Twente), Vikash Kumar (Meta AI Research)