MyoChallenge 2023: Towards Human-Level Dexterity and Agility

Vittorio Caggiano, Guillaume Durandau, HUIYI WANG, Chun Kwang Tan, Pierre Schumacher, Huawei Wang, Alberto Silvio Chiappa, Alessandro Marin Vargas, Alexander Mathis, Jungdam Won, JUNGNAM PARK, Gunwoo Park, Beomsoo Shin, Minseung Kim, SEUNGBUM KOO, Zhuo Yang, Wei Dang, Heng Cai, Jianfei Song, Seungmoon Song et al. (2 additional authors not shown)

Keywords: Reinforcement learning, Neuromusculoskeletal control, manipulation, locomotion

TL;DR: Summary and analysis of a competition on training agents to control musculoskeletal models for dexterous manipulation and agile locomotion

Abstract: Humans move nimbly and with ease, capable of effortlessly grasping items of many shapes and qualities. Over millions of years, the musculoskeletal structure, central and peripheral neural systems have evolved together to provide this capacity. Understanding the underlying mechanisms of this complex system helps translate benefits to other fields, from robot locomotion to rehabilitation. To illicit new insights into the generation of diverse movements and precise control as well as foster collaboration between the biomechanics and the ML community, the MyoChallenge at the NeurIPS 2023 Competition featured two tracks: Manipulation and Locomotion. Manipulation involved precisely manoeuvering an object of varying shape by controlling a 63-musculoskeletal arm model and generating stable grasps. Locomotion involved the combination of abstract reasoning and low-level control, as agents have to chase or evade from a moving object by controlling an 80-musculoskeletal model of human legs. These tasks best highlighted our overarching theme of dexterity and agility, requiring the generation of skilled and efficient movements with realistic human limbs. The Myosuite framework enabled the challenge through a realistic, contact-rich and computation-efficient virtual neuromusculoskeletal model of the human arm and legs. This was the second iteration of the MyoChallenge with 59 teams participating, and over 500 submissions. Each task involved two phases, increasing in difficulty over time. While many teams achieved high performance in phase 1 for the Manipulation track, locomotion showed variable performance across participants. In phase two, scores for all teams dropped significantly as the focus shifted towards generalization under uncertain conditions, highlighting the need for stronger generalization in agents In future challenges, we will continue to pursue the generalizability in dexterous manipulation and agile locomotion, which is crucial for understanding motor constructs in humans.

NeurIPS 23 Competition Analysis Paper: NeurIPS23 Competition Analysis Paper

Submission Number: 2489