RobotisSW/GR00T-N1.7-SH5-Task000905-50k-Hand_Intern

GR00T N1.7 SH5 Task000905 — 50,000 steps

Complete resumable checkpoint at repository root.

Task: Use the right hand to pick up a white sanitary pad from the box and hold it. 51 recorded episodes, 46 train / 5 held out, 30 Hz. Head camera only (cam_left_head), 54 measured joint positions and 90 raw tactile values. Outputs: 40 x 54 absolute joint positions in radians, ordered left arm7, right arm7, left hand20, right hand20. Head/lift/base are excluded from outputs.

Batch64, accumulation1, LR1e-4, warmup1000, BF16/TF32. Frozen VLM backbone; action/projectors and finger_mlp tactile encoder are trained. See artifact_manifest.json for the source commit, normalization and contract. This requires the compatible custom SH5 tactile GR00T implementation, not vanilla GR00T without the tactile adapter. No robot success/safety claim is made.

When ready, model/processor/statistics/embodiment mapping and optimizer, scheduler, RNG, trainer step are included. Raw videos, demonstrations, credentials, server paths, private logs and TrainingArguments pickle are excluded. Verify SHA256SUMS after downloading every file at the recorded commit.

Resume: download this repository completely to checkpoint-50000 inside the original run, retain the original local dataset and launcher, and use train_200k.sh --resume. Restore the saved optimizer/scheduler/RNG and continue the original 200000-update cosine schedule. Offline inference uses Gr00tPolicy with embodiment new_embodiment and use_tactile=True.

Training status

Training stopped at step50000. Last logged train loss1.4492 at step49990; logged gradient norm became infinite from step49890. Tactile gradient was zero in those final logs. This checkpoint is preserved for diagnosis/recovery and is not a validated deployment model. No claim of successful tactile learning is made.