GR00T names a family of models that has changed a great deal since March 2025. The first, GR00T N1, was described in a March 2025 paper and built on NVIDIA’s Eagle-2 vision-language model. The GitHub repository keeps separate branches for N1.5 and N1.6. The current release, N1.7, was announced as an early-access preview on April 17 and is now marked as a general-availability release in the repository.
What changed in N1.7
The repository’s change log lists differences large enough to make old and new results hard to compare:
- The vision-language backbone moved from Eagle to Cosmos-Reason2-2B, built on the Qwen3-VL architecture.
- State and action dimensions grew from 29 to 132, and the action horizon from 16 to 40 steps.
- Pretraining added about 20,000 hours of human video.
- Several model defaults changed, including select_layer, from 16 to 12.
- The command-line flag --action-horizon was renamed --execution-horizon.
The Hugging Face announcement calls the upgrade from N1.6 “a drop-in swap” for existing configurations. The model underneath still changed, so a team comparing a new result with an N1.6 baseline should rerun the baseline.
Where the documents disagree
NVIDIA’s descriptions are not fully consistent. The Hugging Face post describes a “32-layer DiT” action head. The repository’s change log says the action head went “from 32 to 16 diffusion layers.” The deployment guide benchmarks a LIBERO checkpoint with 32 DiT blocks. The checkpoint name varies too: nvidia/GR00T-N1.7 in the Hugging Face post, nvidia/GR00T-N1.7-3B in the repository and developer blog, while NVIDIA’s product pages drop the N and say GR00T 1.7. The model card lists the same Bridge dataset as post-training data for both its Bridge and its Fractal checkpoints.
Performance claims differ in tone. The README says N1.7 “delivers comparable performance to N1.6.” NVIDIA’s July developer blog reports gains on specific benchmarks, including +61% on DROID-F6 and +2% on SimplerEnv Fractal. Both can be true at once, but anyone citing either should name the benchmark.
What to record
Some settings travel inside checkpoints. The repository notes that a padding-mask option stays off for released checkpoints and, when off, can shift an action by about 2 percent of its magnitude depending on the other prompts in a batch. A reproducible result should list the checkpoint name and code commit, the embodiment tag and modality config, the training steps and batch size, and the evaluation trial count. If the hardware or camera setup changed, say so.
NVIDIA’s October 21 release added reinforcement-learning workflows for refining GR00T 1.7 models, so fine-tuned variants will multiply. Results reported without the base version and recipe will get harder to interpret.