Groot-n17-PaperTowelRoll-Tasks726-729-RightArmOnly-Filtered

Full resumable training checkpoint at optimizer step 4,000. The original cosine scheduler horizon is 20,000 steps.

Inputs: cam_left_head, cam_left_wrist, cam_right_wrist; arm_left and arm_right state. Output: arm_right only, 24-step action horizon. Effective global batch size: 96.

This checkpoint was produced by a recovery run from nvidia/GR00T-N1.7-3B with the same filtered Tasks 726–729 dataset and original hyperparameters. The preceding run terminated at step 3,836 during an attempted live stop-control change, before its first scheduled checkpoint; those unsaved states were not recovered.

The model weights, both native DeepSpeed ZeRO-2 optimizer shards, scheduler, trainer state, and both rank RNG states are included. Native optimizer-state resume requires 2 GPU ranks. See resume/RESUME.md for download, environment setup, dataset transfer, and launch instructions.

The training dataset is not included in this public model repository.