CSI analysis

Control subspace inactivation (CSI) — the paper's muscle-synergy analysis. Muscle activations produced by a trained policy are collected, Principal Component Analysis (PCA) is run on them, and the control signal is then projected onto the subspace spanned by the N most important principal components while task performance is measured. The result is a functional measure of how many synergies the policy actually needs.

PCA is run two ways, and comparing them is the point of the analysis:

• Per task — principal components computed from a single task's activations.
• Pooled — principal components computed from all tasks' activations combined.

If the two curves coincide, the policy's synergies transfer across tasks; if the pooled subspace needs many more components to reach the same performance, the low-dimensional structure is task-specific.

CSI analysis vs. CSI-Finetuning

This page projects the actions of an already-trained policy after the fact. It is a different experiment from CSI-Finetuning, where a policy's action space is constrained to a fixed subspace and then trained inside it.

Why only MyoHand tasks

The analysis runs over the 11 MyoHand tasks — the five finger reaches, pen, reorient, and the four Baoding variants. Only these share the same embodiment, and therefore the same 39-muscle action space, so principal components are comparable across them and can be pooled into a shared subspace.

elbow_pose (MyoElbow, 6 muscles), relocate (MyoArm, 63) and kinesis (MyoLeg, 80) have different action dimensionalities and are excluded. See Tasks.

To generate new inactivation results:

1. Collect activations and actions from a trained policy.
2. Fit PCA or NMF components and evaluate inactivation performance.
3. Plot the resulting inactivation curves.

Cumulative explained variance uses the recordings from the first step and can be plotted independently of the inactivation analysis.

Collecting activations and actions

Rolls out a trained policy and saves per-episode observations, action means, rewards, and (for Arnold) intermediate activations. The action means feed the PCA steps below.

• Script: plotting/collect_activations.py
• Output: HDF5 (.h5) files under data/activations/<policy_id>/, e.g. data/activations/example_64670238/

# Collect activations for multiple tasks
for task in hand_thumb_reach hand_index_reach hand_middle_reach hand_ring_reach hand_little_reach reorient pen baoding_p1_ccw baoding_p1_cw baoding_p2 baoding_p2_overlap; do
    python plotting/collect_activations.py \
        --load data/final_benchmarks/example_checkpoint/rl_model_64670238_steps.zip \
        --task $task \
        --num_episodes 100 \
        --arnold \
        --normalize \
        --device cpu \
        --out_dir data/activations
done

Running PCA or NMF inactivation analysis

Supply the checkpoint and the directory containing either per-episode recordings or compact signals:

python plotting/analyze_pca_inactivation.py \
    --load path/to/rl_model_64670238_steps.zip \
    --signals data/activations/example_64670238 \
    --method pca --scope task --num_episodes 100 \
    --out_dir data/pca_analysis/example_64670238/task
python plotting/analyze_pca_inactivation.py \
    --load path/to/rl_model_64670238_steps.zip \
    --signals data/activations/example_64670238 \
    --method pca --scope global --num_episodes 100 \
    --out_dir data/pca_analysis/example_64670238/global

Each run writes curves.csv, episode outcomes and fitted components. PCA centers the recorded action means, then projects stochastic policy actions through the fitted basis. The same global fit is reused across tasks. Performance divides solved steps by the fixed task horizon, including episodes that terminate early.

For the additional NMF analysis, collect physical signals with collect_signals.py, then use --method nmf --n_fits 10. NMF projects physical actuator controls after actuator processing; it does not factor signed policy actions. NMF uses ranks 1–38 and ten seeded fits; PCA uses ranks 1–39. --dimensions selects a subset.

Plotting PCA inactivation performance

python plotting/plot_pca_inactivation.py \
    --curves data/pca_analysis/example_64670238/task/curves.csv \
             data/pca_analysis/example_64670238/global/curves.csv

Without --curves, the command plots the included historical summary at data/analysis/historical_curves.csv. That cache contains 20 episodes per rank for task-specific PCA, 100 for global PCA, and 100 across ten NMF fits. Fresh runs default to 100 episodes, matching the supplied manuscript's PCA methods. Cached results and fresh reproduction results should therefore be distinguished.

Figures are saved under data/figures/pca_inactivation/ as PNG and SVG.

Plotting cumulative explained variance of actions

Plots the cumulative explained variance of the actions (per-task and global) versus the number of principal components.

• Script: plotting/plot_action_pca_variance.py
• Output: data/figures/cumulative_variance/<policy_id>/ (.png / .svg)

python plotting/plot_action_pca_variance.py \
    --activations_dir data/activations/example_64670238 \
    --out_dir data/figures/cumulative_variance/example_64670238

Comparing subspaces

python plotting/analyze_subspaces.py --data_dir data/analysis/signals
python plotting/analyze_subspaces.py --selection capacity_recordings --successful 100

Selections are explicit in data/analysis/reproduction/signals.json; customize them with --selections. Collect all listed policies/tasks first (see Analysis signals). The commands write pairwise measurements, summaries and figures to data/figures/subspaces/. PVD is directed from the first condition to the second; PAD is the mean principal angle in degrees. Shading is standard deviation across pairs. The default selections compare Arnold, expert, multi-task OBC and single-task OBC; these are additional comparisons and are not the multiple-Arnold-checkpoint comparison described in Figure 7 of the supplied PDF.

For learning curves of policies trained inside a constrained subspace, see CSI-Finetuning.