Gom-sy/cafe_stretch_groot_idm_labeled

Cafe Stretch — GR00T v2 corpus (mixed GT / IDM labels)

Hello Robot Stretch serve task (cup → coffee machine), 656 episodes / 1,180,922 frames @ 6 fps, in GR00T v2 (LeRobot-compatible) format. Labels are mixed by domain — the exact recipe we use for policy training:

• Original episodes carry ground-truth (GT) actions (real teleop commands, full amplitude).
• Augmented episodes carry IDM-inferred actions — an Inverse Dynamics Model trained on the GT episodes re-labels the visually augmented videos, enabling policy training (e.g. GR00T N1.6) on augmented data that has no native action labels.

Composition (656 episodes)

Range	Episodes	Domain	Video	Labels
0–39	40	Original	320×240 real capture (bloodmoon3929/ai_challenge)	GT
40–79	40	Original + illumination relight (stratified EV −2…+1)	320×240	GT
80–367	288	Appearance augmentation (SAM-based recolor)	224×224	IDM-inferred
368–655	288	Appearance + illumination augmentation	224×224	IDM-inferred

All augmented episodes reuse the original robot trajectories — only appearance/illumination differ. IDM labeling overwrites the 6 manipulation dims; base dims and progress are preserved from source.

Format

data/chunk-000/episode_XXXXXX.parquet   # observation.state, action (10-dim), timestamps, ...
videos/chunk-000/observation.images.gripper/episode_XXXXXX.mp4   # wrist camera
videos/chunk-000/observation.images.head/episode_XXXXXX.mp4      # head camera
meta/  # info.json, episodes.jsonl, tasks.jsonl, modality.json, stats.json

• Cameras (dual): gripper (wrist) + head. Mixed resolutions are intended — square/resize at the transform level (e.g. crop-scale 0.95 → resize 224×224).
• Action (10-dim raw): base (3, zeros in this task) · lift · arm · wrist roll/pitch/yaw · gripper · progress.
• stats.json: computed over this exact mixed corpus with arm clipped to [0, 0.52] (telescopic limit; negative arm commands are physical-limit violations in the source teleop).

IDM label quality (vs ground truth)

Open-loop evaluation of the labeling IDM (per-dim Pearson corr, physical units):

Group	Original frames	Augmented frames
lift	0.999	0.965
arm	0.990	0.929
wrist	0.995	0.973
gripper	0.992	0.809
macro	0.994	0.919

Known limitation: on augmented frames the (continuous) gripper shows amplitude compression (regression toward the mean). This is why original episodes keep GT labels — they anchor full-amplitude gripper behavior during policy training.

Normalization used in our training

lift/arm: q99 · wrist/gripper: μ-law (μ=3) over q99 anchors ("Mu_law_q99"). Normalization modes are not stored in checkpoints — configure them in your data config.

Provenance

• Source videos: bloodmoon3929/ai_challenge (LeRobot v1) → converted to GR00T v2 (dual-cam gate, depth excluded).
• Augmentation: SAM-based appearance recolor + illumination relight (stratified EV −2…+1).
• Labels: GT from source teleop (original episodes); IDM (GR00T-Dreams IDM head, vision tower frozen, trained on the 80 GT episodes, 15 epochs, global batch 64, final loss ≈ 0.01) for augmented episodes.