Hand smoothness and Baoding PCA¶
Use the Arnold environment. The analyses default to compact recordings
in data/analysis/signals/. To generate additional recordings, use
Analysis signals. The default selections require 100 episodes for each listed policy/task.
Both compact task.h5 and existing task_episode_N.h5 recordings are supported.
Run both analyses from the repository root:
Use --data_dir <signals_directory> and --out_dir <results_directory> to change
the paths. See Analysis signals to collect new recordings.
Smoothness writes per-episode metrics, per-task means, Wilcoxon/Holm comparisons
and smoothness.svg to data/figures/smoothness/. Derivatives use one simulation
step; comparisons pair the 11 task means, with Holm correction within each metric.
Baoding PCA writes dimensionality counts, cumulative variance curves, the human
comparison and baoding_p1_ccw.svg to data/figures/baoding_pca/. PCA centers raw
joint coordinates. Velocities are finite differences within each episode.
Dimensionality is (PCs at 85% + PCs at 95%) / 2, separately for position and velocity.
The human values in data/analysis/reproduction/baoding_human.csv use the Ball rows,
Angle scaling, 20 joints in Tables 1–2 of
Todorov and Ghahramani (2004).
The supplied manuscript PDF does not describe these smoothness metrics or the human Baoding comparison. Smoothness derivatives use a unit timestep, so their values are expressed per simulation step rather than per second.