MyoChallenge 2022: Learning contact-rich manipulation using a musculoskeletal hand

Vittorio Caggiano*, Guillaume Durandau*, Huawei Wang, Alberto Chiappa, Alexander Mathis, Pablo Tano, Nisheet Patel, Alexandre Pouget, Pierre Schumacher, Georg Martius, Daniel F.B. Haeufle, Yiran Geng, Boshi An, Yifan Zhong, Jiaming Ji, Yuanpei Chen, Hao Dong, Yaodong Yang, Rahul Siripurapu, Luis EduardoFerro Diez, Michael Kopp, Vihang Patil, Sepp Hochreiter, Yuval Tassa, Josh Merel, Randy Schultheis, Seungmoon Song, Massimo Sartori, Vikash Kumar

*Corresponding author for this work

Research output: Contribution to journalConference articleAcademicpeer-review

11 Downloads (Pure)


Manual dexterity has been considered one of the critical components for human evolution. The ability to perform movements as simple as holding and rotating an object in the hand without dropping it needs the coordination of more than 35 muscles which act synergistically or antagonistically on multiple joints. This complexity in control is markedly different from typical pre-specified movements or torque based controls used in robotics. In the MyoChallenge at the NeurIPS 2022 competition track, we challenged the community to develop controllers for a realistic hand to solve a series of dexterous manipulation tasks. The MyoSuite framework was used to train and test controllers on realistic, contact rich and computation efficient virtual neuromusculoskeletal model of the hand and wrist. Two tasks were proposed: a die re-orientation and a boading ball (rotation of two spheres respect to each other) tasks. More than 40 teams participated to the challenge and submitted more than 340 solutions. The challenge was split in two phases. In the first phase, where a limited set of objectives and randomization were proposed, teams managed to achieve high performance, in particular in the boading-ball task. In the second phase as the focus shifted towards generalization of task solutions to extensive variations of object and task properties, teams saw significant performance drop. This shows that there is still a large gap in developing agents capable of generalizable skilled manipulation. In future challenges, we will continue pursuing the generalizability both in skills and agility of the tasks exploring additional realistic neuromusculoskeletal models.

Original languageEnglish
Pages (from-to)233-250
Number of pages18
JournalProceedings of Machine Learning Research
Publication statusPublished - 2023
Event36th Annual Conference on Neural Information Processing Systems, NeurIPS 2022 - New Orleans, United States
Duration: 28 Nov 20229 Dec 2022
Conference number: 36


  • hand and wrist
  • manipulation
  • Neuromusculoskeletal control
  • Reinforcement learning
  • 2023 OA procedure


Dive into the research topics of 'MyoChallenge 2022: Learning contact-rich manipulation using a musculoskeletal hand'. Together they form a unique fingerprint.

Cite this