This is the code repository for winning the Boading ball task of the MyoChallenge 2022. Our team was named stiff fingers, and this name was inspired by the fact as the fingers were not too agile.... perhaps bending the fingers themselves is not too task relevant. If you want to read more about the challenge and our solution, check out this exciting article summarizing the challenge and winning solutions in the Proceedings of Machine Learning Research!
We also wrote a manuscript further detailing our solution, and analyzing the learned policy. Analyzing the policy against data from human subjects revealed a number of insights into efficient control of complex biological systems. Check out the preprint: Acquiring musculoskeletal skills with curriculum-based reinforcement learning. You can find code for this analysis here. The relevant data for models and rollouts is available on zenodo.
Our winning team comprised:
- Alberto Chiappa and Alexander Mathis EPFL, Switzerland
- Pablo Tano, Nisheet Patel, Alexandre Pouget University of Geneva, Switzerland
Here we have documented a summary of our approach including our key insights and intuitions along with all the training steps and hyperparameters here.
Listed in requirements.txt. Note that there is a version error with some packages, e.g. stable_baselines3, requiring later versions of gym which myosuite is incompatible with. If your package manager automatically updates gym, do a pip install gym==0.13.0 (or equivanlent with your package manager) at the end and this should work fine. If you experience any issues, feel free to open an issue on this repository or contact us via email.
Run python src/main_baoding.py to start a training. Note that this starts training from one of the pre-trained models in our curriculum. You can find all the trained models along with the scripts used to train them and the environment configurations here. The full information about the training process can be found in the summary.
To evaluate the best single policy network (see the summary), run python src/main_eval.py. To evaluate the final ensemble (55% score), run python src/eval_mixture_of_ensembles.py.
If you use our code, or ideas please cite:
@InProceedings{pmlr-v220-caggiano22a,
title = {MyoChallenge 2022: Learning contact-rich manipulation using a musculoskeletal hand},
author = {Caggiano, Vittorio and Durandau, Guillaume and Wang, Huwawei and Chiappa, Alberto and Mathis, Alexander and Tano, Pablo and Patel, Nisheet and Pouget, Alexandre and Schumacher, Pierre and Martius, Georg and Haeufle, Daniel and Geng, Yiran and An, Boshi and Zhong, Yifan and Ji, Jiaming and Chen, Yuanpei and Dong, Hao and Yang, Yaodong and Siripurapu, Rahul and Ferro Diez, Luis Eduardo and Kopp, Michael and Patil, Vihang and Hochreiter, Sepp and Tassa, Yuval and Merel, Josh and Schultheis, Randy and Song, Seungmoon and Sartori, Massimo and Kumar, Vikash},
booktitle = {Proceedings of the NeurIPS 2022 Competitions Track},
pages = {233--250},
year = {2022},
editor = {Ciccone, Marco and Stolovitzky, Gustavo and Albrecht, Jacob},
volume = {220},
series = {Proceedings of Machine Learning Research},
month = {28 Nov--09 Dec},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v220/caggiano22a/caggiano22a.pdf},
url = {https://proceedings.mlr.press/v220/caggiano22a.html},
abstract = {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.}
}
@article {Chiappa2024skills,
author = {Alberto Silvio Chiappa and Pablo Tano and Nisheet Patel and Abigail Ingster and Alexandre Pouget and Alexander Mathis},
title = {Acquiring musculoskeletal skills with curriculum-based reinforcement learning},
elocation-id = {2024.01.24.577123},
year = {2024},
doi = {10.1101/2024.01.24.577123},
publisher = {Cold Spring Harbor Laboratory},
abstract = {Efficient, physiologically-detailed musculoskeletal simulators and powerful learning algorithms provide new computational tools to tackle the grand challenge of understanding biological motor control. Our winning solution for the first NeurIPS MyoChallenge leverages an approach mirroring human learning and showcases reinforcement and curriculum learning as mechanisms to find motor control policies in complex object manipulation tasks. Analyzing the policy against data from human subjects reveals insights into efficient control of complex biological systems. Overall, our work highlights the new possibilities emerging at the interface of musculoskeletal physics engines, reinforcement learning and neuroscience.Competing Interest StatementThe authors have declared no competing interest.},
URL = {https://www.biorxiv.org/content/early/2024/01/25/2024.01.24.577123},
eprint = {https://www.biorxiv.org/content/early/2024/01/25/2024.01.24.577123.full.pdf},
journal = {bioRxiv}
}
Acknowledgments & funding: We thank members of the Mathis Group for helpful feedback. A.M. is appreciative to the Kavli Institute for Theoretical Physics (KITP) in Santa Barbara, where part of the manuscript was written. AM thanks Nicola Hodges and John Krakauer for discussions on skill learning. A.C. and A.M. are funded by Swiss SNF grant (310030_212516). A.I. acknowledges EPFL's Summer in the Lab fellowship to join the Mathis Group. A.M. was supported in part by grants NSF PHY-1748958 and PHY-2309135 to the KITP. P.T. and N.P. were supported by University of Geneva internal funding.