m3/groot-n1.7-cube-bowl-100steps

GR00T N1.7 — Fine-tuned on cube_to_bowl_5 (SO-101 Pick-and-Place, 100 steps)

This is a smoke-test fine-tune of nvidia/GR00T-N1.7-3B on the cube_to_bowl_5 demo dataset bundled with Isaac-GR00T.

Purpose

Validates the GR00T N1.7 post-training pipeline on a single Blackwell GPU (RTX PRO 6000) under CUDA 12.8. The 100-step run is a pipeline integrity check, not a converged model — production fine-tunes typically run 2,000+ steps.

Training

Parameter	Value
Base model	nvidia/GR00T-N1.7-3B
Dataset	demo_data/cube_to_bowl_5 (5 episodes, ~4150 frames, SO-101 follower arm)
Embodiment tag	NEW_EMBODIMENT
Steps	100 (MAX_STEPS=100)
Global batch size	8
Learning rate	1e-4 (cosine, 5% warmup)
Weight decay	1e-5
Train runtime	155 s on a single RTX PRO 6000 Blackwell
Loss trajectory	1.146 → 0.984
GPU	NVIDIA RTX PRO 6000 Blackwell Server Edition (sm_120, 96 GB)
CUDA / driver	12.8 / 580.126.09
Trainable params	1.62 B / 3.14 B (51.5%)

Files

Inference-only artifacts. The training-only optimizer.pt (~13 GB) and rng_state.pth have been omitted to keep the repository small.

Loading

from gr00t.policy.gr00t_policy import Gr00tPolicy
from gr00t.data.embodiment_tags import EmbodimentTag

policy = Gr00tPolicy(
    model_path="m3/groot-n1.7-cube-bowl-100steps",
    embodiment_tag=EmbodimentTag.NEW_EMBODIMENT,
    modality_config=...,
    modality_transform=...,
    device="cuda:0",
)

Limitations

• Smoke-test only — 100 steps is far below convergence.
• The cube_to_bowl_5 dataset has only 5 episodes; the model is heavily underfit and will not generalize beyond its training distribution.
• For real downstream use, run the full fine-tune at MAX_STEPS=2000+ per the Isaac-GR00T finetune guide.

Citation

The architecture is described in the GR00T N1 white paper.