Analysis signals

plotting/collect_activations.py produces intermediate transformer activations and per-episode HDF5 files under data/activations/. Analysis scripts can read those files directly when their configured policy directory matches the recording folder. The analysis scripts default to the supplied compact files under data/analysis/signals/<policy_id>/<task>.h5. To analyze per-episode recordings, pass --data_dir data/activations and use matching policy selections.

Collect compact hand recordings

Use checkpoints from the current data layout, including their neighboring args.json, vocabulary.json and normalization checkpoint. Supply the actual model path:

python plotting/collect_signals.py \
    --load path/to/rl_model_64670238_steps.zip \
    --task baoding_p1_ccw --arnold --normalize \
    --num_episodes 100 --seed 0 --device cpu \
    --policy_id arnold/seed_0
python plotting/collect_signals.py --expert --task baoding_p1_ccw \
    --deterministic --num_episodes 100 --seed 0 --device cpu --policy_id expert

Repeat for every policy/task in data/analysis/reproduction/signals.json. That file specifies analysis cohorts, not checkpoint download paths. Use each analysis's --selections option to provide a JSON with your local policy IDs. For per-episode recordings use --data_dir data/activations and map policy IDs to the existing recording subdirectories.

Each compact file has episode_N groups with actions, action_means, joint_positions, rewards, and solved. Joint/muscle names specify column order; metadata records seed, fixed horizon and timestep. Actions and positions precede the environment step. Add --physical_signals for muscle_controls and muscle_activations measured after the step. These are distinct from policy actions. Add --num_success 100 --max_attempts 1000 to retain successful episodes only. Failed attempts retain their episode IDs in the seed sequence.

This command supports the 11 MyoHand tasks with the current compositional policy. For locomotion use collect_gait.py.

Export existing recordings

python src/analysis/export_signals.py \
    --input_dir data/activations/example_64670238 \
    --output data/activations/arnold/seed_0/baoding_p1_ccw.h5 \
    --task baoding_p1_ccw --policy_id arnold/seed_0 --horizon 200 \
    --num_episodes 100 --signals actions action_means joint_positions

Set --horizon and --dt to the recording's environment settings. --episodes selects specific episode IDs. The importer reads raw joint positions from the per-episode observation layout and flattens singleton action dimensions. muscle_controls can be exported only when the original file actually contains them (or a physical activations dataset); they cannot be reconstructed from policy actions alone. Physical controls imported from per-episode recordings are marked before_step, following their original recording convention.

The supplied compact recordings can be analyzed directly. Collect or import recordings when adding policies or tasks beyond the supplied selections.

Export obtains joint and muscle column names from recording metadata, or from the current task model when older recordings omit them. For recordings from a different model, supply --axes path/to/axes.json containing joint_names and muscle_names. Analysis can read existing per-episode recordings directly; conversion is optional.