Announcing MyoChallenge ‘23

We are excited to be launching MyoChallenge 2023 for this year’s NeurIPS competition track! In this post, we will briefly outline some of the high-level goals of this challenge, as well as the details of this year’s dual manipulation and locomotion tracks.

Last year, we released MyoSuite, a suite of musculoskeletal models and tasks. This year, we have added two new musculoskeletal models to the suite: MyoLegs (80 muscles, 16 DoF) and MyoArm (63 muscles, 27 DoF). Learning control policies using these two newest MyoSuite models will be the focus of this year’s MyoChallenge.

Motivation

The purpose of MyoChallenge is to advance understandings of neuromuscular control, using MyoSuite’s high-dimensional, physiologically accurate musculoskeletal models as a testbed. The MyoChallenge 2023 seeks to (A) push the boundaries of physiological motor control in complex manipulation and locomotion tasks, while also (B) fostering interdisciplinary engagement from across the biomechanics, motor control, robotics, and machine learning communities. Given the many uniquely complex facets of motor control—including multiarticularity, overactuation, very high-dimensional action spaces, and indirect joint control via pull-only muscle forces—the 2023 MyoChallenge presents among the most challenging control tasks to date. As such, building out solutions to these control problems will represent significant and concrete advances both in reinforcement learning techniques and in our understanding of how biological systems may naturally achieve such impressive neuromotor control.

Track One: Object Manipulation and Placement with MyoArm

The Manipulation Track aims to yield adept single arm-hand dexterity. Participants are tasked with training a policy for the MyoArm that competently grasps, manipulates, and places parametric geometries and common household items to specific orientations and positions within the given workspace.

Participants’ solutions will be evaluated based on the success rate in accomplishing specified goals and energy efficiency. The task is split into two phases. The first phase assesses a policy’s capability in picking up a specific object, manipulating it to achieve randomized orientation and position goals, and placing the object into a desired location. The second phase will further intensify this track by introducing new objects with differing geometries, while also varying physical properties and goal positions.

Track Two: Bipedal Agility with MyoLegs

The Locomotion Track replicates the high-intensity environment of the World Chase Tag competition. The task is to efficiently navigate an environment to either pursue or evade an opponent within a 20-second round.

The task’s evaluation is also multifaceted, incentivizing single policies that can both effectively chase and tag opponents quickly, and successfully evade opponents. The task is also divided into two phases. Phase One focuses on training locomotion policies on a plain Quad with a dynamic opponent, while Phase Two introduces the dual challenge of chasing and evading the opponent, with the addition of obstacles and terrain variations to test the model’s adaptability.

Prize Pool

We’re pleased to announce a generous total prize pool of $20k for this year’s MyoChallenge. Not only will we crown the winners for both the Manipulation and Locomotion Tracks, but we are also introducing the following awards for this year’s competition:

• Community Contribution Award: Awarded to the team that provides significant insights in their method, fostering a richer and more collaborative neuromotor learning community.
• Best Student Teams Award: Given to the student team that showcases the most promising work throughout the competition.
• Innovation Award: Honors the team that exhibits the most inventive and creative approach in solving the competition tasks.

Final Thoughts

We are very excited to be building on last year’s MyoChallenge, which saw incredible participation from 40 teams from over 10 countries submitting more than 340 solutions. We strongly encourage participants to navigate to our MyoChallenge 2023 webpage, where they can find tutorials, starter code, and more technical information about this year’s challenge. Best of luck to all participants!