Here we provide the weights of the neural network policies used for the analysis presented in our article.

The archives whose names start with a number (01 - 32) correspond to the 32 curriculum steps to train the Baoding Balls policy which ranked first at the MyoChallenge 2022. The code used for the training and which can be used to test the policies can be found at https://github.com/amathislab/myochallenge.

The archives hand_pose, hand_reach, pen and reorient correspond to the other policies used in the article. They were developed in the paper Latent exploration for reinforcement learning, Chiappa et al., NeurIPS 2023. They can be loaded and tested with the code at https://github.com/amathislab/lattice.

The archive datasets includes three subfolders: rollouts, umap and csi.

The files in rollouts are the datasets of transitions resulting from the interaction between a policy and the environment. 
The files in umap are the pre-computed projections of specific subsets fo the datasets included in rollouts using UMAP.
The files in csi report the performance of the policies described in our paper when applying Control Subspace Inactivation (CSI).

These datasets are necessary to run the notebooks to reproduce the paper's figures and main results, with the code at https://github.com/amathislab/MyoChallengeAnalysis

If you find these weights useful, please cite: @article{chiappa2024acquiring, title={Acquiring musculoskeletal skills with curriculum-based reinforcement learning}, author={Chiappa, Alberto Silvio and Tano, Pablo and Patel, Nisheet and Ingster, Abigail and Pouget, Alexandre and Mathis, Alexander}, journal={bioRxiv}, pages={2024--01}, year={2024}, publisher={Cold Spring Harbor Laboratory} }

@article{chiappa2024latent, title={Latent exploration for reinforcement learning}, author={Chiappa, Alberto Silvio and Marin Vargas, Alessandro and Huang, Ann and Mathis, Alexander}, journal={Advances in Neural Information Processing Systems}, volume={36}, year={2024} }