Figure R.1. Grouping matrix P_n derived from varying sample sizes, and the distance (Frobenius norm) between them and a 500K-sample grouping matrix. p_{ij} is the probability that muscle i and muscle j are in the same group (averaged over 10 seeds). The grouping results are shown to converge to their final grouping with a data point quantity as low as 25,600. Even if we have only 100 data points, the grouping result is similar to the final result.
Figure R.2. The distance (Frobenius norm) between grouping matrices derived from varying sample sizes and a 500K-sample grouping matrix.
Figure R.3. Learning curves on the Legs-Walk environment. Performance comparison between randomly generating 40 clusters, randomly segmenting consecutive muscles into 40 clusters and our method generating the dynamical synergistic representation of 40 clusters. Mean ± SD across 5 random seeds.
Figure R.4. The original order (muscles are sorted alphabetically in the model by name) of the grouping matrix and visualization of muscles.
Figure R.5. Learning curves in the Baoding and Pen environment from Myosuite. Mean ± SD across 3 random seeds for all the environments.