GR00T N1.7 DROID reproducibility canary

This is a one-optimizer-step fine-tuning canary for NVIDIA Isaac GR00T N1.7. It demonstrates that the pinned source, models, three-episode DROID sample, and Reproducible AI workflow can produce and verify a readable checkpoint. It is not a quality, convergence, or deployment-safety claim.

Reproducible inputs

• Source: reproducible-ai/Isaac-GR00T@ed97646fe7cf438a49af3530571567024beb2edd
• Base model: nvidia/GR00T-N1.7-3B@2fc962b973bccdd5d8ce4f67cc63b264d6886495
• VLM backbone: nvidia/Cosmos-Reason2-2B@9ce19a195e423419c349abfc86fd07178b230561
• Dataset: lerobot/droid_1.0.1@0eabc778f959c54b8c5aa3626cc1128d2d2e54d4 (three episodes)
• Training: one optimizer step, global batch size 1
• Release path: droid-canary-0.0.1

The checkpoint's evaluation.json records the step count, training log, byte size, SHA-256 digest, tensor count, and first tensor shape for each weight shard.

License and safety

The checkpoint is distributed under the NVIDIA License copied into LICENSE. That license restricts this work and derivatives to non-commercial research or evaluation. The upstream GR00T and Cosmos notices are retained under upstream/; Cosmos attribution and its separate license are included in NOTICE and NVIDIA_OPEN_MODEL_LICENSE.md.

This model is not tested or intended for mission-critical applications that require functional safety. Users are responsible for use-case-specific testing, guardrails, and compliance before deployment.