Embedded Files
The following speakers tentatively confirmed their participation in the workshop.
Dagmar Sternad
Dagmar Sternad
Northeastern University, USA
Quantitative Assessments of Functional Skill in Individuals with Neurological Injury
Abstract: Assessments of sensorimotor impairment in persons with neurological injury such as stroke still remain either qualitative, using clinical scales such as the Fugl-Meyer Score, or remain limited to simple ‘abstracted’ movements, such as maximum grip force or horizontal reaching using robotic interfaces. However, for moderately impaired individuals, physical and occupational therapy has long emphasized the value of assessing and training sensorimotor skills that reflect objectives and challenges of daily actions, rather than abstractions from them. Hence, there is a need for more real-life inspired experimental paradigms that leverage modern measurement and modeling advances. With this goal, our lab has developed an experimental testbed motivated by the self-feeding action of guiding a cup of coffee to one’s mouth. The additional theoretical motivation for this task that interacting with a dynamically complex object is an interesting and unresolved control problem that may also reveal or amplify sensorimotor impairments. To facilitate testing stroke patients, we developed a portable set-up that allows video-based extraction of manual interactions with a cup and a rolling ball inside. The custom 3D-printed object complied with a cart-pendulum model of the task that afforded evaluation not only with descriptive kinematics, but also enabled to reconstruct user forces and quantitative modeling of control strategies.
Georg Martius
Georg Martius
University of Tübingen and Max Planck Institute for Intelligent Systems, Germany
Learning to control musculoskeletal models with reinforcement learning
Abstract: Humans excel at robust bipedal walking in complex natural environments. In each step, they adequately tune the interaction of biomechanical muscle dynamics and neuronal signals to be robust against uncertainties in ground conditions. However, it is still not fully understood how the nervous system resolves the musculoskeletal redundancy to solve the multi-objective control problem considering stability, robustness, and energy efficiency. In computer simulations, energy minimization has been shown to be a successful optimization target, reproducing natural walking with trajectory optimization or reflex-based control methods. However, these methods focus on particular motions at a time and the resulting controllers are limited when compensating for perturbations. In robotics, reinforcement learning~(RL) methods recently achieved highly stable (and efficient) locomotion on quadruped systems, but the generation of human-like walking with bipedal biomechanical models has required extensive use of expert data sets. This strong reliance on demonstrations often results in brittle policies and limits the application to new behaviors, especially considering the potential variety of movements for high-dimensional musculoskeletal models in 3D. I will present our results on achieving natural locomotion with RL without sacrificing its incredible robustness. We believe this will have big impacts on future healthcare applications.
Hao Su
Hao Su
New York University, USA
Bridging Sim2real Gap for Autonomous Control of Exoskeletons via Learning-in-Simulation and High-Torque Motors
Abstract: Can we design wearable robots for everyone and everywhere? This talk introduces a new design paradigm that leverages custom high-torque density motors to electrify robotic actuation. This allows our wearable robots to achieve exceptional performance, including ultra-compact and lightweight exoskeletons/exosuits, along with high compliance and bandwidth in human-robot interactions. The presentation will also cover a data-driven, physics-informed reinforcement learning framework that accelerates control policy development in simulation, significantly reducing wearable robot development time. Our learning-in-simulation controllers bridge the sim-to-real gap and reduce energy consumption during activities like walking, running, and stair climbing, leading to significant energy savings for users. Additionally, our advances in bionic limbs enhance mobility and manipulation for individuals with musculoskeletal and neurological injuries. We envision these innovations sparking a paradigm shift in wearable robotics, transforming them from lab-bound rehabilitation devices to ubiquitous personal robots for everyone, everywhere - in applications such as workplace injury prevention, pediatric and elderly rehabilitation, home care, and sports.
Vittorio Caggiano
Vittorio Caggiano
FAIR-MetaAI, USA
From Muscle Synergies to Dexterous Control: Advancing Motor Learning with AI and MyoSuite
Abstract: Human movement relies on complex and continuous sensory-motor integration, enabling quick adaptation to changing environments and the acquisition of new skills. How humans control movements is still an open question in neurology and neuroscience. A key challenge lies in the limited ability to access comprehensive, real-time data from both central (brain/spinal cord) and peripheral (muscles/tendons) nervous systems during movement. This constraint makes direct observation difficult, driving the need for simulations to study and replicate human motor control.
Existing models of the peripheral nervous system (muscle, articulations) typically focus on bio-mechanical details with limited capabilities to represent the dynamic interactions between the body and its environment. For instance, conventional simulators fail to adequately capture the discontinuous hand-object interactions crucial for dexterous manipulation. Furthermore, typical biomechanics simulators computational inefficiency prevents them from fully leveraging modern data-driven methods. These challenges motivated the development of MyoSuite, a novel framework designed to model skilled, dexterous human behavior with physiological accuracy, real-time environmental interactions, and compatibility with current machine-learning pipelines.
In this presentation, I will introduce our work on MyoSuite and highlight two new studies that leveraged this platform to study physiological sensory-motor dexterous and agile behaviors. In the first study, MyoDex, I will be showing how, by using human-like complex multi-task learning, we can create representations that generalize across tasks. This generalization is supported by the emergence of coordinated muscular activation patterns activated as a unit: muscle synergies. In the second study, by using a representation based on muscle synergies (Synergistic Action Representation, SAR), I will show how it is possible to easily generalize to new tasks. Using this approach, AI agents can manipulate hundreds of distinct objects robustly, outperforming SOTA by approximately 40% while simultaneously being up to 4x more efficient in learning new contact-rich dexterous manipulation tasks! Beyond manipulation, the same approach can also be applied to locomotion.
These results demonstrate that AI-driven approaches, when aligned with human physiological constraints, can uncover the same physiological priors humans rely on for motor control. This has far-reaching implications for creating physiological digital twins capable of capturing individual motor control capabilities.
Shaoping Bai
Shaoping Bai
Aalborg University, Denmark
Modeling, sensing and control of human-robot interaction in wearable exoskeletons
Abstract: Wearable exoskeleton technology is being advanced rapidly for broad applications. An essential issue in exoskeleton development is the human-robot interaction, which requires comprehensive study in modeling, interaction sensing and control at kinematic, physiological and cognitive levels. This talk will provide a brief overview of wearable technology development at the Exoskeleton Lab, Aalborg University, addressing research challenges in human-exoskeleton interaction. Novel design and sensing methods will be presented, along with application examples.
Yanan Sui
Yanan Sui
Tsinghua University, China
Self Model for Embodied Intelligence: Building Neural Representation from the Bottom Up
Abstract: Consciousness refers to the ability of individuals to be aware of their own state, behavior, and emotion. The self-model for embodied intelligence plays an important role in understanding human behavior and self-consciousness. This work constructed a comprehensive model of the human musculoskeletal system and part of the peripheral motor nervous system. Through online optimization and reinforcement learning methods in high-dimensional spaces, we can simulate the perception and control of the human body's neuro-muscular-skeletal dynamics process. The self-model of embodied intelligence may help us understand the neural control of movement from the bottom up.
Atsushi Takagi
Atsushi Takagi
NTT Communication Science Laboratories, Japan
Why humans make errors
Abstract: Unlike robots, humans are unable to reproduce the same movement precisely. The consensus in neuroscience is that this movement variability comes from noise in the muscles, which scales with activation size. But an increasing body of evidence suggests that something is amiss with this explanation. In this talk, I propose that movement variability is primarily due to mistimed muscle activity. I will present several behavioral experiments to support this hypothesis, and show how computational modelling of mistimed muscles may even explain the coordination problems caused by degenerative diseases that damage the brain such as Huntington's and Parkinson's disease.
Page updated
Google Sites
Report abuse