UsanoCoCr/ProAct-Data

ProAct data exports

Data exports located alongside the trained ProAct motion checkpoints. Source code: Social_robot_baseline / opensource. Paper: ProAct.

Archive	Contents	Scope
processed_seg.tar.gz	Original processed SeG feature directory: training/validation/test NPZ exports, per-clip motion/text features, scalers and motion preprocessing pipelines	The processed SeG data is publicly redistributed with the project owner's explicit authorization. Original SeG motion capture files are not included.
photoreal_retargeted.tar.gz	30 scenes; 30 retargeted BVH files, 60 left/right WAV files, 30 state TXT files	Photoreal-derived subset in the historical photoreal_bvh_retarget2noitom_new export, with a Noitom-format retargeted skeleton. This is not the complete audio training mixture.

manifest.json records every archived relative path, byte size, SHA-256, archive hash, and count. Archives retain the original exported files without refitting statistics or regenerating splits. Tar ownership and timestamps are normalized.

Download and use

pip install huggingface_hub
hf download UsanoCoCr/ProAct-Data --repo-type dataset --local-dir data/ProAct-Data
cd data/ProAct-Data
shasum -a 256 -c SHA256SUMS
tar -xzf processed_seg.tar.gz
tar -xzf photoreal_retargeted.tar.gz
export MOTION_CONTROLNET_DATA_ROOT="$PWD/processed_seg"
export PROACT_TEXT_DATA_PIPE="$MOTION_CONTROLNET_DATA_ROOT/data_pipe_30fps.sav"
export PROACT_TEXT_OUTPUT_SCALER="$MOTION_CONTROLNET_DATA_ROOT/output_scaler.sav"

The SeG export contains 6,550 files (6,546 NPZ and 4 SAV). Root training arrays are float32 with input shape [3, 8633, 77, 512] and output shape [3, 8633, 150, 31]; validation has 11 windows and each of the two named test exports has one. The joint_rot directory contains 3,165 augmented SeG motion clips and their mirrored copies, across 208 action classes. These names match the historical SeG mix_bvh export; no Photoreal/ZEGGS scene names were found in this subset. Per-clip joint_rot arrays have 31 channels; text arrays hold three 77-token, 512-dimensional conditions and their names. All NPZ member headers were checked: numeric/Unicode arrays only, with no object dtype. These are historical export sizes, not a claim that every count matches the paper's final dataset recipe. One per-clip text filename retains a comma in THUMB,_FOREFINGER_AND_LITTLE_FINGER_RAISE, whereas its motion name omits it; account for this historical naming difference when rebuilding per-clip data. The original files have been preserved.

The SeG export is the text-data root associated with the published epoch-800 ControlNet. Its top-level input_scaler.sav, output_scaler.sav, and data_pipe_30fps.sav match the original checkpoint validation assets. Use the source project's PyMO compatibility loader for historical .sav exports. NPZ files contain precomputed arrays; SAV files are Python/joblib serialization used by the source pipeline.

The retargeted Photoreal BVHs are intermediate motion assets. They are not ready-to-send robot commands or the standardized 31-dimensional model tensors. Apply the source project's preprocessing and configure the corresponding paths before training.

Provenance and exclusions

The Photoreal subset contains exactly scenes 01–30 from the local Photoreal export. All 60 audio files and 30 state files were checked byte-for-byte against its historical photoreal_pxb source directory. The project mappings connect these scene names to PXB184. This release does not claim independent checksum authentication against the official upstream archive.

The complete historical audio feature directory also includes renamed ZEGGS scenes. It is withheld because permission to redistribute ZEGGS-derived data has not been obtained. The audio-domain scalers and pipelines fitted to that mixture are not included. Download original sources from audio2photoreal and ZeroEGGS, observing their terms. The supplied Photoreal subset does not reproduce those mixed training splits or statistics.

Original SeG data remains available through the official Semantic Gesticulator release and its agreement. ProActBench is not included. These exports and checkpoints do not establish exact reproduction of paper metrics.

Terms and citation

See TERMS.md: the two subsets retain different source terms. The source code's MIT license does not apply to these datasets. Use the data for non-commercial research and cite ProAct, Semantic Gesticulator, and From Audio to Photoreal Embodiment.