Scientists have created an innovative AI-based method that greatly enhances our comprehension of the intricate motor functions of the human hand. The research team implemented a unique machine learning technique that integrated curriculum-based reinforcement learning with precise biomechanical models. Their study offers a comprehensive, dynamic, and anatomically correct representation of hand movements, drawing directly from how people learn complex motor skills.
In the fields of neuroscience and biomedical engineering, accurately modeling the sophisticated motions of the human hand has been a longstanding challenge. Existing models often fail to effectively portray the intricate relationship between the brain’s motor commands and the corresponding actions of muscles and tendons. This gap not only obstructs scientific advancement but also restricts the creation of effective neuroprosthetics for individuals with limb loss or paralysis.
Professor Alexander Mathis from EPFL and his team have pioneered an AI-based technique that significantly deepens our understanding of these motor functions. They devised an inventive machine learning method that merged curriculum-based reinforcement learning with intricate biomechanical simulations.
Mathis’s study introduces a thorough, dynamic, and anatomically precise model of hand movement, directly influenced by how humans acquire complex motor skills. This research not only secured the MyoChallenge award at the 2022 NeurIPS conference but has also been published in the journal Neuron.
Virtually controlling Baoding balls
“What excites me most about this research is that we’re exploring the essential principles of human motor control — a long-standing enigma. We’re not merely constructing models; we’re revealing the fundamental mechanics of how the brain and muscles collaborate,” Mathis expresses.
The NeurIPS challenge organized by Meta urged the EPFL team to explore a fresh method within the AI realm known as reinforcement learning. The challenge entailed creating an AI that could precisely manipulate two Baoding balls, each governed by 39 muscles in a highly coordinated fashion. Although the task might seem simple, it is incredibly complex to emulate virtually due to the sophisticated dynamics of hand movements, which necessitate muscle synchronization and balance maintenance.
In this intensely competitive setting, three graduate students — Alberto Chiappa from Mathis’s team, Pablo Tano, and Nisheet Patel from Alexandre Pouget’s team at the University of Geneva — significantly outperformed other entrants. Their AI model achieved a flawless success rate in the initial phase of the competition, outshining the nearest competitor. Furthermore, during the more challenging second phase, their model demonstrated exceptional resilience and maintained a substantial lead, ultimately winning the competition.
Breaking tasks down into smaller segments — and repetition
“To succeed, we were inspired by the way humans learn complex skills through a method known as part-to-whole training popular in sports science,” explains Mathis. This part-to-whole training influenced the curriculum learning technique employed in the AI model, systematically decomposing the intricate task of controlling hand movements into smaller, achievable components.
“To overcome the constraints of existing machine learning models, we utilized a method called curriculum learning. After 32 distinct stages and nearly 400 hours of training, we successfully taught a neural network to accurately control a realistic representation of the human hand,” Chiappa remarks.
A crucial factor contributing to the model’s success is its capacity to identify and leverage basic, repeatable movement patterns, termed motor primitives. Interestingly, this method of learning behavior may provide insights into neuroscience, specifically how the brain influences the acquisition of motor primitives for mastering new tasks. The complex interaction between brain functions and muscle control highlights the challenges involved in creating machines and prosthetics that genuinely emulate human movement.
“Achieving a range of movements and developing a model akin to the human brain is essential for performing various daily tasks. Each task, while potentially divisible into smaller parts, requires distinct sets of these motor primitives to ensure effective performance,” Mathis explains.
Utilizing AI to explore and comprehend biological systems Silvestro Micera, a prominent researcher in neuroprosthetics at EPFL’s Neuro X Institute and a collaborator of Mathis, emphasizes the vital significance of this research for future advancements and understanding the limitations of current advanced prosthetics. “We currently lack a deeper insight into how we achieve finger movement and grasping control. This research is precisely aligned with addressing this crucial aspect,” Micera observes. “We recognize the importance of integrating prosthetics with the nervous system, and this study provides a robust scientific foundation that fortifies our approach.”
