Meta’s Toddler-Like AI Agents Act As Biomechanical Models That Learn Like Humans, Help in Developing More Realistic Robots and Avatars

In a simulated environment, disembodied body parts powered by artificial intelligence explore the basic motion of a human body. This is part of the latest innovation from the MyoSuite AI platform, where machine learning is applied to biomechanical models.

Simulation for Musculoskeletal Motor Control

Artificial intelligence company Meta AI released the MyoSuite 2.0 collection in collaboration with researchers from McGill University in Canada, Northeastern University in the U.S., and the University of Twente in the Netherlands. The project aims to apply machine learning in demonstrating dexterity and agility to human-like models. The platform offers a collection of baseline musculoskeletal models and open-source standard tasks for experts to explore.

According to Vikash Kumar, one of the project's lead researchers, duplicating the human body's motor strategies in MyoSuite is more difficult than moving a robot around. However, he is certain that robot developers can learn valuable insights from the control techniques of the human body.

The cerebral Fundamental AI Research (FAIR) branch of Meta AI initiated the project, but experts are still unsure how this innovation can be applied to the commercial products of Meta. When MyoSuite version 1.0 was released in May 2022, Mark Zuckerberg announced that this research could help develop more realistic avatars for the metaverse.

Last year, the researchers ran a contest called MyoChallenge 2022, where the 40 participating teams were asked to control a simulated hand to rotate a die. Although the participants successfully trained their algorithms to accomplish the tasks, they were found to be weak at generalizing. If the properties or locations of the objects are changed, the algorithms will find it hard to accomplish simple tasks.

Based on that weakness, the Meta team developed new AI agents better at generalizing from one task to another using learning platforms MyoArm and MyoLegs. The research team thought of switching from training an algorithm to find solutions to specific tasks to teaching representations to help it look for solutions.

Associate Professor Emo Todorov from the University of Washington believes that the focus of MyoSuite on learning general representations indicates that control strategies can be helpful for a whole set of tasks. He also added that the generalized control strategies of the platform are similar to the neuroscience principles of muscle synergies.

Role of Biomechanical Models in Studying Human Motion

Biomechanical models refer to the mathematical representations used in conducting qualitative or quantitative analyses of the complex interactions of human musculoskeletal systems. They are a description of the human body as a mechanical device.

In recent years, virtual humans have been used in assessing different products in various industries, such as aerospace, automotive, task simulation, and personnel training. The human body is characterized by an articulated system where internal forces produce joint movements in the body segments. In this sense, biomechanical models are important tools for studying human motion. They allow researchers to study problems that cannot be analyzed directly on humans or have too expensive experimental costs.

Check out more news and information on Artificial Intelligence in Science Times.