easyminnn/utt-tactile-prepared

UTT tactile prepared data

Preprocessed tactile datasets for pretraining the Unified Tactile Tokenizer (UTT): a sensor-agnostic tactile encoder/decoder (code: branch tactile-tower-tva of the cosmos-framework fork, cosmos_framework/model/generator/tokenizers/utt/). Nothing here is new data: every folder is a re-encoding of a public dataset. Each dataset keeps its original license — read its SOURCE.md (origin, citation, license, exact transformations) before use. Several sources are non-commercial (e.g. CC BY-NC 4.0); research use only.

Layout

v0/
  stats_v0.json                      per-sensor-family normalization stats used by v0 training
  pretrain/<dataset>/                S0 pretraining data
    <dataset>-NNNN.tar               ~2 GB uncompressed tar shards (payload is JPEG / npz)
    MANIFEST.json                    sha256 + size per shard
    SOURCE.md                        origin, citation, license, transformations
  downstream/<dataset>/              benchmark simulations kept OUT of pretraining
                                     (RoboTwin Clean50 tactile, UniVTAC, Bench2Dex TacMap)

Pretrain datasets (v0): digit_force, feelanyforce, fota, hpt, tacquad, egotactile, egotouch, sparsh_skin. Images follow the Sparsh convention (320x240 H x W, landscape frames rotated 90 deg clockwise, background frame stored per sequence); arrays/F/T are raw physical values.

Use

git clone <cosmos-framework fork> && cd cosmos-framework && git checkout tactile-tower-tva
pip install -r scripts/utt/requirements.txt
DATA_ROOT=/data/utt bash scripts/utt/run_s0.sh setup   # downloads + verifies sha256 + unpacks this repo
DATA_ROOT=/data/utt bash scripts/utt/run_s0.sh check   # 20-step smoke
DATA_ROOT=/data/utt bash scripts/utt/run_s0.sh s0a     # pretraining stage S0-a

Versions: a new preprocessing format is published as a new top-level folder (v1/, ...); v0/ is never modified in place.