In a cutting-edge digital simulation, a skeletal arm, driven by artificial intelligence, meticulously examined a miniature toy elephant, flipping it with precision. Using 39 muscles spread across 29 joints, this disembodied arm mirrored the innocent exploration of a toddler. Moving on, it manipulated objects ranging from toothpaste tubes and staplers to alarm clocks. In parallel, skeletal legs, equipped with 80 muscles and 16 joints, performed toddler-like kicks and movements.

This display of AI-operated body parts is the jewel of the MyoSuite platform. Unveiled today by tech giant Meta AI, in collaboration with prominent universities including McGill, Northeastern, and the University of Twente (Netherlands), MyoSuite 2.0 is an audacious attempt to merge machine learning with biomechanical complexities.

With these creations, the team aspires to emulate the dexterity and agility of human movement, a feat made challenging given the intricate coordination of large and small muscle groups. This release also offers an assortment of baseline musculoskeletal models and benchmarks, aiming to empower fellow researchers.

Mark Zuckerberg, Meta’s iconic leader, hints at a broader vision. He says, “This research could also help us develop more realistic avatars for the metaverse.”

Vikash Kumar, a central figure behind the research, underscores the complexity. Unlike robots that employ a single motor for each joint, the human anatomy boasts multiple muscles per joint, and these muscles traverse various joints. This presents a significant challenge as even a simple motion requires continuous muscular engagement. Kumar, transitioning from a dual role at Meta and Carnegie Mellon to focus on CMU’s Robotics Institute, posits that human body control techniques may harbor lessons for robotics. “Evolution has chosen this intricate form for a reason,” he emphasizes.

Though conceptualized by Meta AI’s cerebral arm, FAIR, it’s not hard to envision the potential commercial applications of MyoSuite.

In the MyoChallenge 2022 contest, participants were tasked with controlling a simulated hand to execute fine motor activities. However, while impressive results emerged, there was a stark lack of generalization in the AI’s performance.

To address this, the Meta team embarked on creating AI agents adept at transitioning skills from one task to another. Using MyoArm and MyoLegs as experimental platforms, the team’s focus shifted from solving specific problems to teaching the AI broader concepts. Kumar describes the process as allowing the AI to play with various objects, just like a toddler. This broad-based approach significantly expedited task-specific learning.

The upcoming MyoChallenge 2023 will test teams on their ability to pick up, manipulate, and place household items using MyoArm and play tag using MyoLegs.

Vittorio Caggiano, a Meta scientist with a dual background in AI and neuroscience, believes that insights from the MyoSuite can propel the fields of neuroscience and biomechanics forward. The crux lies in understanding and then translating these insights across domains.

Experts outside the direct Meta sphere echo this optimism. Emo Todorov, an associate professor at the University of Washington, acknowledges the strides made by MyoSuite, emphasizing the importance of the platform’s generalized control strategies. Drawing parallels with neuroscience, Todorov elucidates how such strategies might mirror the human brain’s approach to muscular movement, adding, “MyoSuite is innovatively reconstructing these principles.”

Yet, as Meta continues its pioneering trajectory, the real challenge might be offering these AI agents a comprehensive physiological understanding. After all, as any parent will attest, true toddler comprehension often requires a more holistic, sometimes even mouthy, exploration of objects.