💻 computer science

MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence

MyoChallenge 2025 is a NeurIPS benchmark that leverages high-fidelity musculoskeletal simulations and machine learning to evaluate and advance motor control intelligence through upper and lower limb sports tasks, uniting a global interdisciplinary community to bridge gaps in robotics, biomechanics, and neuroscience.

Original authors: 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, B

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, high-stakes video game tournament, but instead of playing Mario Kart or FIFA, the contestants are trying to teach a digital robot how to play sports just like a human athlete. This is the story of MyoChallenge 2025.

Here is the breakdown of what happened, using simple analogies.

The Big Problem: The "Muscle" Gap

Think of current robots like cars with very stiff, simple engines. They can move, but they don't have "muscles." Real humans have thousands of tiny muscles that pull on bones, stretch, and contract in complex ways to make us agile.

Scientists have been trying to build "Digital Twins" (perfect computer copies) of human bodies that have these realistic muscles. But until now, it's been like trying to teach a toddler to play professional tennis using a toy car. The computer models were too simple, and the physics engines were too slow to handle the complexity of real human movement.

The Solution: The "Olympics" for Digital Muscles

To fix this, a group of researchers created MyoChallenge 2025. Think of this as the Olympics for Artificial Intelligence, but the athletes are computer programs, and the events are:

Table Tennis: A robot arm and torso must hit a ping-pong ball back and forth.
Soccer Penalty Kick: A full-body robot must run up and kick a soccer ball into a goal (sometimes with a goalie trying to stop it).

The goal wasn't just to make the robot move; it was to make it move with human-like agility and muscle coordination.

The Rules of the Game

The competition used a special video game engine called MyoSuite. This engine is like a super-advanced physics simulator.

The Bodies: The robots weren't simple blocks. They were detailed models with 273 muscles for the table tennis player and 290 muscles for the soccer player.
The Challenge: The robots had to learn by trial and error. They couldn't just be told "move your leg this way." They had to figure out how to coordinate hundreds of muscles simultaneously to balance, run, and strike a ball.

The Two Events

1. The Table Tennis Rally (The Upper Body)

The Setup: A robot arm and torso holding a paddle.
The Task: A ball comes flying in at different speeds and angles. The robot must hit it back perfectly.
The Difficulty: It's not just about hitting the ball; it's about holding the paddle steady, moving the torso for balance, and reacting instantly.
The Winner (ActingAI): They didn't just let the robot guess. They gave it a "coach" (a physics planner) that predicted where the ball would go. Then, they used a clever trick to simplify the math, turning the complex "muscle" problem into a simpler "joint" problem, making it much easier to learn.

2. The Soccer Penalty Kick (The Full Body)

The Setup: A full robot body (legs and torso, no arms) facing a soccer ball.
The Task: Run up to the ball and kick it into the net, even if a goalie is there.
The Difficulty: This is incredibly hard. The robot has to balance on one leg, swing the other, and time the kick perfectly, all while managing 290 muscles.
The Winner (Servette MyoClub): They didn't start from scratch. They took a robot that already knew how to walk (like a human baby learning to walk) and gave it extra training to learn how to kick. It was like taking a professional runner and teaching them how to shoot a soccer ball, rather than trying to teach a newborn to do both at once.

The Results: Who Won?

Participation: Almost 70 teams from around the world joined, including students, doctors, and AI experts.
The Scores:
- In Table Tennis, the winning team hit the ball back 100% of the time in the easy round and 64% in the hard round.
- In Soccer, the winning team scored 32% of the time in the hard round.
The Gap: While the robots did well, they still aren't quite as good as elite human athletes. The paper notes that the robots lack the "explosive power" and perfect timing that real humans have, mostly because the computer simulations are still a bit slow and clunky compared to real life.

The Big Takeaway

The paper claims that this competition proved we can finally simulate complex sports with realistic muscles. However, it also highlighted a major bottleneck: computing power.

The Analogy: Training these robots is like trying to run a super-complex simulation on a slow, old laptop. The winning soccer team took 14 days to train their robot.
The Future: The authors say the next step is to move these simulations to super-fast graphics cards (GPUs). This would be like upgrading from a bicycle to a race car, allowing researchers to train these digital athletes much faster and more realistically.

In a Nutshell

MyoChallenge 2025 was a global contest to see if AI could learn to play sports using realistic human muscles. The winners showed that by combining smart math tricks and using pre-trained "walking" skills, robots can learn to play ping-pong and soccer. But to make them truly human-level athletes, we need faster computers to handle the massive amount of muscle math required.

Technical Summary: MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence

Problem Statement

Human athletic performance represents the pinnacle of sensorimotor intelligence, requiring rapid decision-making, precise control, and coordinated execution within dynamically changing environments. Replicating this capability in artificial intelligence and robotics remains elusive because current frameworks often rely on simplified actuator and dynamics models that fail to capture the biological complexity of muscle-based systems (nonlinear, delayed, and coupled characteristics). Furthermore, understanding biological mastery is hindered by the difficulty of measuring complex muscle coordination in vivo. While previous research has focused on isolated kinetic performance or partial body segments, there is a critical lack of public benchmarks that integrate high-fidelity, full-body musculoskeletal (MSk) models with interactive, physically simulated sports environments and advanced machine learning algorithms.

Methodology

MyoChallenge 2025 addresses this gap by establishing a benchmark within the MyoSuite open-source framework, powered by the MuJoCo physics engine. The competition features two distinct tracks designed to test motor control intelligence in sports:

Table Tennis Track: An agent controls a biomechanical upper limb (MyoArm: 27 DOFs, 63 muscles) and trunk (MyoTorso: 18 DOFs, 210 actuators) to return an incoming ball. The pelvis is actuated in x/y to simulate lateral movement. The task requires a generalist policy capable of handling varying ball trajectories and speeds.
Soccer Penalty Track: An agent controls a full-body model (including MyoLeg: 28 DOFs, 80 muscles, and MyoTorso) to execute a penalty kick against a goalkeeper. This task demands coordinating 290 muscles to simultaneously manage locomotion, postural stability, and a precise ballistic strike.

Simulation and Evaluation:

Models: The MSk models are nearly three times more complex than previous iterations, featuring high-fidelity muscle-tendon units.
Phases: The competition consisted of an opening phase (static environments) and a playoff phase (introducing significant environmental variations in mass, friction, ball trajectories, and goalkeeper behaviors) to test robustness.
Metrics: Teams were ranked primarily by success rate (successful returns or goals). In the playoff phase, ties were broken by "muscle activation effort" (mean square activation), rewarding energy efficiency.
Data: No pre-collected training datasets were required; policies were expected to learn via interaction. However, a custom bot was provided to generate imitation learning data, and public datasets (AMASS, BONES-SEED) were encouraged as priors.

Key Contributions and Winning Solutions

The competition attracted 69 teams and 562 submissions. The winning solutions demonstrated a shift from purely end-to-end learning toward hybrid approaches incorporating physics-based priors and hierarchical structures.

Table Tennis Track

1st Place (Team ActingAI): Proposed Diff-Muscle, a hierarchical framework leveraging differential flatness. They proved that the system's states and controls can be algebraically determined by joint configurations (flat outputs). This allowed them to compress the action space from high-dimensional muscle activations ( $R^{273}$ ) to a lower-dimensional joint space ( $R^{32}$ ). The framework uses a physics-based planner to predict ball trajectories and compute target racket states, a PPO-based RL policy for joint control, and a Kinematics-based Muscle Actuation Controller (K-MAC) to map joints back to muscles.
2nd Place (Team BioSyn): Utilized a structured reinforcement learning approach inspired by the human motor cortex. They decomposed control into "synergy subnetworks" dedicated to specific body regions (hand, arm, torso, pelvis). Exploration was guided by multivariate Gaussian noise weighted by these synergies, reducing antagonistic co-activation and improving energy efficiency.
3rd Place (Team LNSGroup): Integrated physical priors with a CrossQ algorithm and a Gaussian Process Classifier (GPC). A physics model predicted ball trajectories to create dense geometric tracking rewards. The GPC dynamically routed incoming balls to specialized sub-policies (forehand/backhand) based on confidence, effectively handling the broadened trajectory distribution of the playoff phase.

Soccer Penalty Track

1st Place (Team Servette MyoClub): Succeeded by leveraging a pre-trained locomotion prior (KINESIS) rather than training from scratch. They employed a two-stage pipeline:
1. Teacher Policy: Fine-tuned the KINESIS motion imitation policy (trained on human locomotion data) into a target-reaching policy and then a ball-kicking policy.
2. Retargeting: Transferred the teacher policy to the MyoChallenge environment via offline trajectory collection, student pre-training (behavioral cloning), and on-policy behavior cloning (OBC/DAgger) to handle the observation space mismatch. Finally, they applied PPO fine-tuning to improve energy efficiency.

Results

Participation: 69 teams from over 10 countries, with 562 total submissions. 63% of participating teams were composed entirely of students.
Table Tennis Performance: The winning team (ActingAI) achieved 100% success in Phase 1 and 64% in Phase 2, with an effort score of 0.1306. The top three teams were separated by only a 5% difference in success rate in the final scoring.
Soccer Performance: Due to the extreme complexity (290 muscles), only one team surpassed the baseline. Servette MyoClub achieved 100% success in Phase 1 and 32% in Phase 2, with an effort of 1.3174.
Computational Bottleneck: Training times highlighted a severe CPU bottleneck. The winning soccer policy required ~14 days of training, and the table tennis policy required ~51 hours, underscoring the need for GPU-accelerated simulation pipelines.

Significance and Claims

The paper claims that MyoChallenge 2025 establishes the first benchmark for sports agility using high-fidelity MSk simulations, bridging the gap between current state-of-the-art models and true human athletic intelligence.

Interdisciplinary Impact: The event united a diverse community of physicians, neuroscientists, and machine learning experts, serving as a reproducible testbed for research in biomechanics, sports science, and neuroscience.
Methodological Shift: The results indicate that for high-dimensional, dynamic tasks, purely end-to-end learning is often insufficient. Successful strategies increasingly rely on physics-based priors, hierarchical planning, muscle synergies, and behavioral cloning from expert policies.
Future Direction: The paper identifies the lack of end-to-end GPU acceleration as a critical barrier. It argues that the field must shift toward fully GPU-accelerated pipelines (e.g., MuJoCo-Warp) to enable scalable, full-body musculoskeletal control, paving the way for a data-driven era in biomechanical simulation.
Limitations: The authors modestly note that a performance gap remains compared to elite human athletes. Current reward formulations maximize forward velocity but do not explicitly capture optimized kinetic chain sequencing or the stretch-shortening cycle, resulting in simulated kicks that lack the peak velocity and flexible precision of human performance.

The competition suite remains publicly available within the MyoSuite repository to facilitate further research and the development of realistic, generalizable digital twins for rehabilitation and assistive technologies.