MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence
Cheryl Wang, Chun Kwang Tan, Balint K. Hodossy, Eric Lyu, Jun Guo, Wentao Zhao, Huaping Liu, Chengkun Li, Merkourios Simos, Bianca Ziliotto, Alexander Mathis, Siyuan Liu, Jiahao Chen, Shanlin Zhong, Bo Jiang, Ci Song, Yaoye Zhu, Chenhui Zuo, Yanan Sui, Mohamed Irfan Refai, Massimo Sartori, Guillaume Durandau, Vikash Kumar, Vittorio Caggiano, [0.6em] Team ActingAI · 2026
Read MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence by Cheryl Wang, Chun Kwang Tan, Balint K. Hodossy, Eric Lyu, Jun Guo, Wentao Zhao, Huaping Liu, Chengkun Li, Merkourios Simos, Bianca Ziliotto, Alexander Mathis, Siyuan Liu, Jiahao Chen, Shanlin Zhong, Bo Jiang, Ci Song, Yaoye Zhu, Chenhui Zuo, Yanan Sui, Mohamed Irfan Refai, Massimo Sartori, Guillaume Durandau, Vikash Kumar, Vittorio Caggiano, [0.6em] Team ActingAI (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-15650.txt.
Introduction
Cheryl Wang and twenty-three colleagues report MyoChallenge 2025, a NeurIPS test bed for athletic motor skill built on lifelike muscle simulation. Two tracks test upper-limb table tennis rallies and lower-limb soccer penalty kicks with true-to-life arm, leg, and trunk models. Almost seventy teams filed over 560 entries. The contest pushed control methods spanning motion planners, behavior cloning, hierarchical planning, and muscle synergies. It leaves a reusable open test bed for machine learning, biomechanics, and sports science.
Catalog record
| CREATOR | Cheryl Wang, Chun Kwang Tan, Balint K. Hodossy, Eric Lyu, Jun Guo, Wentao Zhao, Huaping Liu, Chengkun Li, Merkourios Simos, Bianca Ziliotto, Alexander Mathis, Siyuan Liu, Jiahao Chen, Shanlin Zhong, Bo Jiang, Ci Song, Yaoye Zhu, Chenhui Zuo, Yanan Sui, Mohamed Irfan Refai, Massimo Sartori, Guillaume Durandau, Vikash Kumar, Vittorio Caggiano, [0.6em] Team ActingAI |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65488 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.15650
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence Cheryl Wang, Chun Kwang Tan, Balint K. Hodossy, Eric Lyu, Jun Guo, Wentao Zhao, Huaping Liu, Chengkun Li, Merkourios Simos, Bianca Ziliotto, Alexander Mathis, Siyuan Liu, Jiahao Chen, Shanlin Zhong, Bo Jiang, Ci Song, Yaoye Zhu, Chenhui Zuo, Yanan Sui, Mohamed Irfan Refai, Massimo Sartori, Guillaume Durandau, Vikash Kumar, Vittorio Caggiano, [0.6em] Team ActingAI Abstract Athletic performance represents the pinnacle of human motor intelligence, demanding rapid choices, precise control, agility, and coordinated physical execution. Replicating this seamless combination of capabilities remains elusive in current artificial intelligence and robotic systems. Concurrently, understanding the biological mastery of these movements is hindered because complex muscle coordination is rarely measured in vivo due to the limitations of physical equipment. To bridge this fundamental gap in understanding, MyoChallenge at NeurIPS 2025 established a pioneering benchmark for motor control intelligence in sports, leveraging high-fidelity musculoskeletal models within physics simulation combined with machine learning-driven algorithms. The competition introduces two distinct tracks emphasizing either upper or lower limbs control: a table tennis rally task utilizing a biomechanic upper limb composed of an arm with a hand (MyoArm) and a trunk (MyoTorso); and a soccer penalty kick using a biomechanic model of legs (MyoLeg) and a trunk (MyoTorso). Marking the fourth iteration of the MyoChallenge series, this event attracted almost 70 teams and over 560 submissions globally, uniting a diverse community ranging from physicians and neuroscientists to machine learning experts. The competition facilitated the development of several state-of-the-art control algorithms for a musculoskeletal system capable of sports agility, leveraging techniques such as physics-based motion planners, on-policy behaviour cloning, hierarchical planning, and muscle synergies. By integrating standardized tasks and physiologically realistic models into the open-source framework of MyoSuite, MyoChallenge’25 serves as a reproducible and reusable testbed to accelerate interdisciplinary research across machine learning, biomechanics, sports science, and neuroscience. Project page: https://www.myosuite.org//myochallenge/myochallenge-2025 1 Introduction Human athletic performance represents one of the most complex manifestations of sensorimotor intelligence, involving continuous control over high-dimensional musculoskeletal (MSk) systems in dynamically changing environments. Understanding human athletic intelligence is a critical pursuit for advancing artificial intelligence systems 29 . In robotics, it is particularly vital for enabling machines to perform agile, adaptive behaviors in unstructured and unpredictable environments. Emulating human-like coordination 31 , agility 17 , and rapid decision-making equips robots to navigate uncertain terrain 66 , collaborate with humans in real time 64 , and respond robustly to unforeseen perturbations 66 . However, reproducing this behaviour in a human-like MSk embodiment is difficult because most of the current RL and robotic control frameworks rely on simplified actuator and dynamics models that do not capture the biological complexity of muscle-based systems, which require nonlinear, delayed and coupled characteristics of muscle-based actuation. Understanding muscle recruitment patterns in athletic intelligence is also crucial in domains like sports and rehabilitation, where biomechanically accurate simulation can provide insights into recovery and performance training. Studying athletic behavior under fast, dynamic, and uncertain conditions offers a unique opportunity to understand how the body integrates perception, prediction, and control—insights that can advance prosthetics, rehabilitation, and assistive tech 53 ; 40 ; 38 ; 41 ; 48 . In addition, these studies can enhance biomechanically grounded animation, virtual athletics, and embodied AI agents in simulation environments 52 . Previous MSk research in sports science has largely focused on isolated kinetic performance 35 or estimating muscle forces during specific athletic tasks 6 ; 15 ; 1 . Furthermore, these forward simulations are generally restricted to partial body segments and rely heavily on tracking captured MoCap data, limiting their ability to simulate generalized, whole-body sports movements. In recent years, the community has seen remarkable progress across biomechanics, machine learning 22 ; 36 ; 43 ; 13 ; 21 ; 12 , assistive technologies ( 2 ; 63 ; 54 ) , and physics-based simulation 16 ; 33 ; 19 ; 9 . Within computational biomechanics, several benchmarks have emerged over the past decade to evaluate assistive devices ( 26 ; 55 ) and general human motor control 25 ; 49 ; 11 ; 10 ; 8 . However, no public benchmark currently integrates high-fidelity, full-body MSk models with interactive, physically simulated sports environments and advanced machine learning algorithms. Developing such a benchmark is essential to quantify the gap between current state-of-the-art models and true human athletic intelligence. To address this critical gap, we introduce MyoChallenge’25, a competition designed to establish the first benchmark for sports agility using MSk simulations. This platform advances the development of realistic biomechanical digital twins of humans, seeking to answer a central question: Can physiological digital twins achieve human-level coordination during complex sports activities? The competition features two independent tracks focused explicitly on upper and lower limb coordination. Crucially, these challenges go far beyond static or repetitive motions. They demand generalization through reactive, adaptive embodied behaviors grounded in the physics of muscle, tendon, and joint dynamics, requiring real-time perception-action loops for agile motor control. Ultimately, this initiative aims to establish a realistic, generalizable, and reproducible benchmark for whole-body control strategies, advancing the state-of-the-art (SOTA) in sports simulation, rehabilitation, and neuromotor research. 2 The MyoChallenge’25 Competition A B Figure 1 : Two tracks of MyoChallenge 2025. A. The Table Tennis track in which an agent is required to hold a paddle and return an incoming table tennis ball B. The Soccer track where the agent needs to score a penalty goal with or without a goalie. For three consecutive years, the MyoSuite team – a multi-univeristy opensource and openscience initiative to advance human musculoskeletal motor control – has hosted the MyoChallenge as an official NeurIPS competition to advance the benchmark for controlling human digital twins using frontier machine learning algorithms. Marking its fourth edition, MyoChallenge’25 presents two tracks focused on athletic intelligence. Compared to previous MyoChallenge editions, the MSk models in 2025 are nearly three times more complex in terms of both the number of muscles and the degrees of freedom, enabling a richer and more challenging investigation of coordinated motor behavior. The first track required the coordination of muscles of the hand, arm and trunk to coordinate the motion and rotation of a paddle to return an incoming table tennis ball. The second track required the coordination of a full-body model, including torso and lower limb muscles, to execute a penalty kick. The competition was divided into an opening phase and a playoff phase. In the initial phase, submissions were evaluated using static environments with no task variations, encouraging participants of all skill levels to test early solutions. Teams scoring above 10% advanced to the playoff phase, which introduced significant environmental variations to test model robustness. The following sections outline the simulation environment and musculoskeletal models (Sec. 2.1 ), the proposed tasks and evaluation criteria (Sec. 2.2 ), and the results of the competition (Sec. 3 ). 2.1 Simulation Environment, Models, and Data Simulation Framework The competition and the complete set of musculoskeletal models are embedded within MyoSuite, an open-source collection of environments and tasks powered by the MuJoCo physics engine ( 51 ) . MyoSuite provides physiologically realistic musculoskeletal models 56 within a simulation framework that significantly surpasses the speed of state-of-the-art simulators 23 ; 18 used in prior competitions, such as