Episodes Preview right_7dof_inspire_f1_6dof Visualizer
100 episodes · 30 fps

TipDex-Data

Hugging Face repository: TipDex-VTLA/TipDex-Data.

This dataset contains 50 original converted robot demonstration episodes and 50 synthetic RGB illumination variants, totaling 100 episodes, 94,172 frames, and 50 independent source recordings. Task: “Move the egg into the box.” Synthetic variants are derived observations paired with the original recorded actions; they are not additional physical demonstrations or proof of successful execution under changed lighting.

Data and pairing

  • LeRobot v2.1 layout; all 100 episodes are in train. Episodes 0–49 are original; episode i+50 is the perturbed counterpart of episode i.
  • Three embedded RGB views (top, left_wrist, right_wrist), each 224×224; no external videos are required. Original encoded image bytes are unchanged. Perturbed RGB is stored as lossless PNG.
  • Action: 13 measured right arm/hand joint coordinates. State: 19 coordinates. Tactile: 364 raw uint16 transport words, not a verified physical taxel/pressure map.
  • All original per-frame observations, actions, flags, tasks, frame indices and timestamps are preserved. Only episode_index and global index are renumbered; only RGB is perturbed in the derived episodes.
  • provenance/pairs.jsonl contains original source basenames, episode IDs, recipes, seeds, decoded RGB hashes and file hashes. meta/episode_sources.jsonl maps every output episode.
  • These 50 recordings are the single-segment sources in the historical training partition. The one split training source and all 8 development / 4 final-evaluation sources are excluded. Do not split paired variants or frames of the same source across train/evaluation.

Synthetic perturbations

Five categories contain 10 derived episodes each: low exposure, strong exposure, directional illumination, moving soft shadow / periodic exposure, and asymmetric camera illumination. Every category has 6 mild (0.25–1 EV magnitude), 3 moderate (1–2.5 EV), and 1 severe (2.5–4 EV) recipes. These are EV parameter-magnitude tiers, not calibrated image-distortion or task-difficulty scores. A fixed recipe and seed are used for each complete episode. Original source timestamps drive smooth temporal changes. Inverse sRGB, linear-domain exposure, clipping, and sRGB encoding follow the copied lighting code. Strong exposure may saturate pixels. Moving shadows are synthetic, not actual human occlusions. Some recipes intentionally preserve one or two camera views.

The lighting implementation is reused from the 2026-11-19 experiment. This release materializes new complete-episode variants with the documented balanced assignment; it is not an export of the old anchor-feature cache. code/build_dataset.py records the deterministic selection and recipe assignment.

Read the data

from pathlib import Path
from datasets import load_dataset

root = Path('/path/to/downloaded/dataset')
ds = load_dataset('parquet', data_files={'train': str(root / 'data/chunk-000/*.parquet')})['train']
image = ds[0]['observation.images.top']  # PIL RGB image
action = ds[0]['action']

Alternatively, download the repository with huggingface_hub.snapshot_download(repo_id="TipDex-VTLA/TipDex-Data", repo_type='dataset') and point a LeRobot v2.1-compatible reader to that root. Read meta/info.json for the original timestamp, action, camera alignment, and tactile validity semantics. FPS is nominally 30; retain the recorded timestamps for temporal analysis. The nine original raw numeric/command/tactile stream sidecars for each selected source are bundled unchanged under meta/raw_streams/{source_stem}/ (450 files, shared by each pair). Raw MCAP files are not bundled.

Validation

provenance/validation.json reports the checks: 100 complete episodes; 50 one-to-one pairs; all 282,516 stored camera images fully decoded; clean image bytes unchanged; lossless perturbed pixels checked against the deterministic transform; all non-RGB payloads identical except the two index fields; no development/final-evaluation source overlap; source parquet SHA256 unchanged. Statistics are recalculated for state, action and RGB; image statistics use every pixel and population standard deviation. provenance/sha256_manifest.json inventories every other release file.

The data has not been newly adjudicated for task success. Preserve the original validity flags. No new license is granted by this packaging operation; consult the data owner for reuse permission.

Downloads last month
63