This is the code repository for winning the Boading ball task of the MyoChallenge 2023. Our team was named CarbonSiliconAI
Our team comprised (Manipulation track):
- Heng Cai, Jianfei Song, Zhuo Yang and Wei Dang, CarbonSiliconAI
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 ofgymwhichmyosuiteis incompatible with. If your package manager automatically updates gym, do apip 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. - download myoarm_object_v0.16(mj237).mjb and put it to
data/myosuite/assets/myo_sim/arm
Run python src/main_relocate.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/eval_sb3.py.
- Caggiano, V., Durandau, G., Wang, H., Chiappa, A., Mathis, A., Tano, P., Patel, N., Pouget, A., Schumacher, P., Martius, G., Haeufle, D., Geng, Y., An, B., Zhong, Y., Ji, J., Chen, Y., Dong, H., Yang, Y., Siripurapu, R., Ferro Diez, L.E., Kopp, M., Patil, V., Hochreiter, S., Tassa, Y., Merel, J., Schultheis, R., Song, S., Sartori, M. & Kumar, V.. (2022). MyoChallenge 2022: Learning contact-rich manipulation using a musculoskeletal hand. Proceedings of the NeurIPS 2022 Competitions Track, in Proceedings of Machine Learning Research 220:233-250 Available from https://proceedings.mlr.press/v220/caggiano22a.html.
- https://github.com/amathislab/myochallenge