UTDSINewsMyoChallenge: Machine Learning Challenge on digital hands

MyoChallenge: Machine Learning Challenge on digital hands

In cooperation with Meta (known for Facebook) AI Lab, UT researchers developed a new open-source software called MyoSuite. Those same researchers now challenge anyone working with AI or Machine learning to use MyoSuite to develop controllers that can solve dexterous manipulation tasks on a realistic digital hand. On behalf of UT, TechMed Centre and DSI sponsor the Machine Learning Challenge called MyoChallenge.

The MyoChallenge consists of two tracks. The first challenge is to make the digital hand reconfigure a die to match desired goal orientations. "Manipulating a single object seems simple enough, however, this task requires delicate coordination of various muscles to move the die to the correct side without dropping it,” says Dr Guillaume Durandau, one of the initiators of the challenge.

In the second track, participants will make the hand rotate two balls around each other. “Simultaneous rotation of two balls is even more complex. This not only requires dexterity like in the first track but also a high level of coordination,” explains Dr. Durandau.

Two stages

The entire competition is split into two stages: an open stage and a play-off stage. In the open stage, anyone can join. Each participant can submit a maximum of five solutions per day. This stage is open right now until the 20th of October. All participants are automatically ranked on a leaderboard and the top scorers can participate in the play-off stage.

In the play-off stage, the participants submit their top-3 policies which will be tested in a new environment. This means that the organizers will test the policies with new physics settings, for example, balls with a higher bounciness. Dr Durandau: “We wanted to really challenge the submitted solutions. Using this setup, we can tune the difficulty of the final stage based on the results of the first stage.”

How to participate in the Challenge?
click here

More information

The play-off stage ends on the 28th of October. Results will be announced during NeurIPS 2022 conference, the main conference on machine learning. More information on the challenge and how to participate can be found on the website of MyoChallenge and updates can be accessed by following the Twitter account of the challenge (@MyoChallenge). The Myochallenge builds upon MyoSuite. MyoSuite supports the co-simulation of AI-powered musculoskeletal systems physically interacting with assistive robots such as exoskeletons.

The challenge is organized by Dr. Guillaume Durandau, Dr. Huawei Wang and Prof. Massimo Sartori from the UT, Dr. Vittorio Caggiano and Dr. Vikash Kumar from Meta AI, Dr. Seungmoon Song from Northeastern University and Dr. Yuval Tassa from Deepmind.

drs. M.M.J. van Hillegersberg - Hofmans (Martine)
Press relations (available Mon, Tue, Thu)