Skip to content
Robotics

GR00T N1.7 для Unitree G1

яблоко на тарелку по команде

2 min read

Это файнтюн nvidia/GR00T-N1.7-3B для embodiment UNITREE_G1_SONIC. Модель управляет Unitree G1 с руками Dex3 через whole-body контроллер NVIDIA GEAR-SONIC v1.1. По камере головы, состоянию робота и текстовой инструкции она выдаёт 40 шагов (0,8 с) motion tokens SONIC и цели для рук Dex3.

Ключевые особенности

  • Файнтюн GR00T N1.7 под тег UNITREE_G1_SONIC, обучен через Isaac-GR00T launch_finetune.py.
  • Вход — камера головы, состояние робота и инструкция; выход — 40 шагов SONIC и цели Dex3.
  • Данные desk_v1: 824 эпизода, 160 415 кадров при 50 Гц. Записи Arena-G1-Static-PickNPlace-Task воспроизведены object-centric через SONIC на симуляционном двойнике стола; яблоко и тарелка в пределах ±8 см, оставлены только успешные rollouts.
  • 20 000 шагов, global batch 16, color jitter, одна A100 40GB, около 3 ч. Это checkpoint-20000.
  • Только симуляция: на реальном роботе и NVIDIA SONIC deploy модель не запускали.

Пример

python gr00t/eval/run_gr00t_server.py --model-path <this repo> --embodiment-tag UNITREE_G1_SONIC --port 5550
ПодробнееHeld-out: 69% (59–77%) на 100 эпизодах — объекты на 12 см вне обучающего диапазона; 23 из 31 ошибки — яблоко сбито до захвата. In-distribution: 95% (91–97%).
Динамика: 99% на скользких поверхностях и лёгком яблоке, часть ячеек ещё не готова. Обучение: одна A100 40GB, ~3 ч.
Ограничения: только симуляция; данные и in-distribution оценка — с одного desk twin. Скорость рук ~1 м/с, рядом с людьми нужен rate-limit.
Discuss Telegram 0 7

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 12 GB · ready files13 GB
  • BF16weights 5.9 GB6.9 GB
  • FP8 / INT8recommendedweights 2.9 GB3.7 GB
  • 4-bitweights 1.6 GB2.3 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F32 — from RTX 5080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # kiteml/gr00t-n1.7-g1-sonic-desk-apple
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download kiteml/gr00t-n1.7-g1-sonic-desk-apple --local-dir ~/models/gr00t-n1.7-g1-sonic-desk-apple
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of GR00T-N1.7-3B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face

About the model

kiteml/gr00t-n1.7-g1-sonic-desk-apple
Author
kiteml
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

70

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments