GR00T N1.7 — Humanoid Benchmark
GR00T N1.7 checkpoints fine-tuned on Humanoid Benchmark tasks (2026-11-11 training round: global batch 696 = 87 per GPU x 8 GPUs, one node per run, 20 epochs).
| Folder | Task | Robot |
|---|---|---|
b6-pedal-bin/ |
Pick up a chip barrel, step on the pedal bin, drop it in | Unitree G1 |
a3-fridge-open/ |
Open the fridge door | Unitree G1 |
a7-cabinet-close/ |
Close the cabinet door | Unitree G1 |
a18-pour-water/ |
Pour water from a pitcher into a glass | H2 (Sharpa hands) |
a10-trolley-push/ |
Push a trolley | Unitree G1 |
a20-basket-transfer/ |
Transfer a basket | Unitree G1 |
a17-unscrew-bottle/ |
Unscrew the cap of a bottle | H2 (Sharpa hands) |
Each task folder has epoch-010/ and epoch-020/ (100 demonstrations per task, 20 epochs, saved at epoch 10 and 20).
Every epoch folder is self-contained for deployment and evaluation: weights, model / processor configuration,
normalization statistics and protocol.json (input/output dimensions, cameras, step counts, file list).
Training records are included for reproducibility (training_plan.json, train.log, and for GR00T trainer_state.json,
training_args.bin, wandb_config.json); optimizer / scheduler / RNG / DeepSpeed resume state is not. code/ holds the adapter code used for training and inference.
Model inputs are identical across all benchmark models: G1 state 44-D (29 body joint angles + 6 + 6 measured hand
values + 3 projected gravity), H2 state 78-D (75 joint angles + 3 projected gravity), no joint velocities.
Training used the Humanoid Benchmark online data augmentation spec v2 (photometric, geometric, sensor and lighting
augmentation plus proprioception noise/delay/drop), with the model's own image augmentation turned off.
The VLM backbone nvidia/Cosmos-Reason2-2B is loaded from its own (gated) repository.
The previous training round of this repository is kept in the git history at revision 2e8f375584631cae0454911e58b5b7850ea9f678 (older batch setting).