Meta Researchers Introduce a New Embodied AI Platform, Called MyoSuite, That Applies Machine Learning (ML) to Biomechanical Control Problems by Unifying Motor and Neural Intelligence

This Article is written as a summay by Marktechpost Staff based on the Research Paper 'MYOSUITE A CONTACT-RICH SIMULATION SUITE FOR MUSCULOSKELETAL MOTOR CONTROL'. All Credit For This Research Goes To The Researchers of This Project. Check out the paper, Github, blog and project.

Please Don't Forget To Join Our ML Subreddit

Meta Researchers introduce a new embodied AI platform called ‘MyoSuite’ that combines motor and neural intelligence to solve biomechanical control problems using machine learning (ML). To meet the data requirements of modern machine learning (ML) algorithms, MyoSuite’s muscle models are up to 4,000 times faster than other simulators.

Since physiologically realistic movements such as twirling a pen or manipulating Baoding balls can be generated, this research could significantly impact areas such as the development of prosthetics and post-injury rehabilitation.

In the metaverse, these models will aid in creating avatars that move more realistically, making the experience more expressive and immersive. ā€”

The more intelligent an organism, the more complex its motor behavior is capable of demonstrating. It’s worth asking, then, what provides the motor control necessary to carry out such difficult decisions. MyoSuite is developed to investigate this question, which includes a set of musculoskeletal models and tasks that can be used to apply ML to biomechanic control issues. MyoSuite brings together the motor and neural aspects of intelligence. MyoSuite and a set of comprehensive benchmarks are being released for the ML community to do further research in this field.

Musculoskeletal biomechanics is multi-joint, multi-actuator physiology. Multiple muscles control each joint’s movement in a single contraction (see image below). A biomechanical system with billions of neurons in the central nervous system and a musculoskeletal system that translates these intentions into actions requires effective coordination between these two systems to achieve intelligent behavior.

Artificial intelligence (AI) aims to mimic intelligent behavior by simulating brain structures or neural networks. The biomechanics community has developed musculoskeletal systems through in-vivo and ex-vivo investigations to comprehend peripheral actuation simultaneously and generally independently.

There are many advantages to MyoSuite’s physiologically realistic musculoskeletal models compared to previous models. This gives us a realistic representation of the human body’s many muscles. Dexterous manipulation, such as spinning a pen or rotating a key, and progress on complex tasks such as simultaneously manipulating two Baoding balls in one hand have all been accomplished, which includes several previously unsolved motor control behaviors. Because now realistic physiological motions can be produced in great detail, the findings are expected to aid in solving complicated real-world issues, including rehabilitation, prosthetics, and ergonomics.

Source: https://tech.fb.com/artificial-intelligence/2022/05/myosuite/

Using ML to control musculoskeletal movement

Data and computation scalability allows for the evolution of solutions that are otherwise infeasible for humans to implement. High-dimensional, complicated problems like Alphago or MuZero, whose closed-form analytic solutions are unknown, can now be solved with modern ML paradigms. However, it is rare for these algorithms to be applied to more sophisticated motor control scenarios, such as musculoskeletal control.

Existing in-silico frameworks like OpenSim, which feature physiologically precise musculoskeletal models, lack the ability to interact with the physical world beyond the agent’s body. To meet the data requirements of machine learning algorithms, these existing frameworks are neither incorporated in sophisticated and skilled motor tasks nor are they computationally effective or scalable. MyoSuite fills these voids.

An ecosystem for musculoskeletal motor control called MyoSuite

MyoSuite was created from the ground up to serve as a comprehensive research platform for studying the physiological mechanisms underlying musculoskeletal motor control. The MyoSuite ecosystem includes a full set of well-tuned, physiologically accurate musculoskeletal models that support temporally interaction-rich musculoskeletal dynamics, and carefully designed behavioral tasks that expose motor control challenges that are aligned with real-world situations, such as daily tasks, injury rehabilitation, and prosthetic/exoskeleton assistance. Video Link

Consequences in the real world

MyoSuite’s comprehensive platform makes it possible to implement real-world applications like rehabilitation, surgery, and shared autonomy assistive devices, as well as synthesize behaviors.

The tendon transfer is a well-known regaining function after a torn tendon. MyoSuite can readily replicate this procedure. By identifying activation patterns and muscle groups, the MyoSuite models can be used to aid design splints and rehabilitation treatments. MyoSuite may also simulate the results of a specific surgical procedure in terms of mobility and its effect on the functional rehabilitation of the damaged areas following the surgical process. – Video Link

To mimic Pytorch, MyoSuite will become the platform of choice for integrating motor and brain intelligence. MyoSuite’s full potential has yet to be realized. MyoSuite aims to become the standard platform for AI-driven solutions that harness musculoskeletal motor control in numerous industries and domains in the future.

Source: https://tech.fb.com/artificial-intelligence/2022/05/myosuite/

MyoSuite has been updated to include accurate human anatomy simulations of the arm and hand. Biomechanics researchers can’t do all of the functions encased in these models. The ML community will have access to new benchmarks that are more detailed than those already available. This allows the community to compare different data-driven solutions using a more extensive collection of reference models.