Learning Physiological Dexterity  |   @NeurIPS 2022 Competition Track  | Sponsorship/Prizes: 75,000+ $ USD

 Status: <open submission>  |   Tutorials   |   Workshop |   Contact-us 



Introducing MyoChallenge - a NeurIPS 2022 competition track on learning contact-rich manipulation skills for a physiologically realistic musculo-skeletal hand. The goal of MyoChallenge is to push our understanding of physiological motor-control responsible for nimble and agile movements of the human body. In the current edition of MyoChallenge, we are focusing on developing controllers for contact rich dexterous manipulation behaviors.  This challenge builds upon the MyoSuite ecosystem -- a fast (>4000x faster) and contact-rich framework for musculoskeletal motor control. 

Competition Tracks
The MyoChallenge consists of two tracks:

Single object track

Task: Die reorientation
Reconfiguring a die to match desired goal orientations. This task require delicate coordination of various muscles to manipulate the die without dropping it.

Double object track 


Task: Baoding balls
Simultaneous rotation of two balls over the palm. This task requires both dexterity and coordination. The objective is to achieve relative rotation of the balls around each other without dropping them.

Competition Stages + Timeline
The entire competition is split into two stages:

This stage is open to everyone. Each participant will be constrained to submit at maximum 5 solutions per day and automatically ranked on a leaderboard. This phase will last about 2 months.

Only the top scorers from the Open Stage will be allowed to participate in the play-off stage. In this case, participants will be asked to submit their top 3 policies which will be evaluated on a new test environment (unknown task physics settings) to the participants. This Play-off Stage will be open for one week after the Open Stage is finished. In addition to preventing overfitting, the two-stage approach will allow us to tune the difficulty of the final stage based on the results of the first stage.

Results will be announced during NeurIPS 2022 conference

2022 Neurips Leaderboard

Phase 1 - Die Reorentation:

Place Team Score

1 pkumarl 0.71

2 IARAI-JKU 0.43

3 AL4Muscles 0.3

Phase 2 - Die Reorentation :

Place Team Score

1 IARAI-JKU 0.13

2 pkumarl 0.12

3 AL4Muscles 0.07

Phase 1 - Baoding Balls:

Place Team Score

1 stiff_fingers 1

2 AL4Muscles 0.98

3 pkumarl 0.62

Phase 2 - Baoding Ball:

Place Team Score

1 stiff_fingers 0.55

2 AL4Muscles 0.41

3 IARAI-JKU 0.15

How to start

Steps:  documentation   tutorials pre-trained behaviors train yourself

MyoChallenge is built upon the MyoSuite ecosystem that consists of a set of physiologically realistic musculoskeletal models and a suite of tasks defined in OpenAI gym interface. The MyoSuite ecosystem contains comprehensive documentation and elaborate tutorials allowing users to gradually familiarize themselves with the platform. MyoSuite hosts a diverse collection of manipulation tasks (right figure), associated pre-trained behaviors policies, and training code to replicate results (which can an be used as launch-pad for MyoChallenge). MyoChallenge simply extends the MyoSuite with the addition of two unsolved tasks -- DieReorientation and Baoding Balls. 

How to Participate

MyoChallenge is hosted on EvalAI. Register your team on the EvalAI platform and head to our submissions page for details on how to submit solutions. 

NeurIPS Workshop

MyoChallenge workshop @ NeurIPS'22 is platform for our community to come together. The organizers will share the analysis and lessons from MyoChallenge'22. The winners will share their winning strategy. Team MyoSuite will be sharing their vision about the platform and next steps. We will also have an exciting line of invited speakers to broaden our horizons! See more details here


We have received generous sponsorship in the form of $65k from Google Cloud Credits to support computing and $10k from the University of Twente (TechMed Centre and Digital Society Institute) for travel grants to the top winners.



 author = {Caggiano, Vittorio AND Wang, Huawei AND Durandau, Guillaume AND Song, Seungmoon AND Tassa, Yuval AND   Sartori, Massimo AND Kumar, Vikash},

 title = {MyoChallenge: Learning contact-rich manipulation using a musculoskeletal hand},

 howpublished = {\url{ }},

 year = {2022}