scilwb/SightTouch-RGBD-Tactile

SightTouch RGB-D and native tactile supervision

Real-robot research data for learning two-finger tactile wrench observations and A/B 6x6 strain-array responses from right-wrist RGB-D images.

Contents and reproducibility

Each Parquet row contains lossless 224x224 RGB uint8 and aligned depth uint16 PNGs, 84 signed native targets, missing-value masks, source camera frame identity, reference-domain identity and timing metadata. Each cohort manifest records input hashes, pixel/target/mask parity verification, file hashes and split counts. The source frame identifiers permit duplicate/leakage auditing across cohorts. RGB-D is the only visual-model input; sensor values are supervision labels. Depth pixels retain original integer counts. Use the cohort's recorded depth_scale_m when present; older online manifests may leave this unknown. Factory depth scale has not been independently metrologically verified.

Offline cohorts preserve declared trial train/validation groups. Online cohorts preserve 20s train / 2s embargo (gap) / 10s later temporal validation blocks. Gap and validation rows must not enter gradient replay. These are development splits, not a blind new-object test. Missing measurements stay masked/null.

Physical and statistical limitations

UMI values are SDK_NATIVE_UNVERIFIED; array targets are signed ADC differences from a captured background, not calibrated Newton force or pressure. Reference sessions can have different zero offsets; do not silently mix reference domains. Fy or other weakly excited channels may be constant. Two fingers are independent observations; equal/opposite numerical predictions are not imposed. Physical contact-frame registration, per-cell Newton calibration and object-side multiaxis force calibration remain unverified. Some online observations are no-contact; frame count does not establish independent grasp/object diversity. Native slip labels are not supplied by the present acquisition API. Object/assembly/placement identity remains unknown where it was not recorded. No synthetic force labels are introduced. Original full-resolution experiments and private provenance are retained separately by the owner; these are processed training-view images.

Use

from datasets import load_dataset
# Select a config listed above, e.g. offline-reference-v1.
data = load_dataset("scilwb/SightTouch-RGBD-Tactile", "offline-reference-v1")

Public/private boundary

The public export uses an explicit field allowlist. Credentials, SSH material, runtime configuration, private logs and vendor SDK binaries/weights are not part of this dataset. Original device identifiers are represented by stable hashes. File content is verified after downloading the pinned Hub commit.

License

The owner has not specified a data license. Public visibility alone does not grant additional usage permissions. No license for third-party SDKs or weights is granted here.