Skip to content
Robotics

LingBot-VLA 2.0 (6B)

RoboTwin/Aloha-AgileX — полный SFT для роборук

1 min readupdated

Полнопараметрический supervised fine-tune robbyant/lingbot-vla-v2-6b от liwy26 на clean-сплите RoboTwin 2.0 Aloha-AgileX (50 задач × 50 демонстраций). 6B: Qwen3-VL-4B backbone + action expert. На входе 3 RGB-вида, 14-D состояние робота и языковая инструкция; на выходе — 50-шаговый action chunk (14-D).

Ключевые особенности

  • Полный SFT, не LoRA: все параметры слиты в 3 safetensors-шарда (BF16, 12.77 ГБ, 1708 тензоров).
  • Три камеры: camera_top, camera_wrist_left, camera_wrist_right; ключи симулятора cam_high, cam_left_wrist, cam_right_wrist.
  • 50-шаговый action chunk, 10 шагов денойзинга, image size 256×256, state/action dim 14.
  • В комплекте нормализация и конфиги: assets/norm_stats/robotwin.json, configs/robot_configs/robotwin.yaml; lingbotvla_cli.yaml — авторитетный inference-конфиг.
  • Не грузить через AutoModel: нужен deploy/ loader; teacher-веса не включены, alignment-тензоры отбрасываются.

Пример

cd /path/to/lingbot-vla-v2-6b-robotwin-aloha-agilex
export QWEN3VL_PATH=/path/to/Qwen3-VL-4B-Instruct
export PYTHONPATH=/path/to/lingbot-vla-v2:$PYTHONPATH
python -m deploy.lingbot_vla_v2_policy --model_path . --use_length 50 --use_bf16 True --use_fp32 False --port 8006
ПодробнееSuccess rate: clean — 50.50%, randomized — 27.40%, overall — 38.95% (779/2000 эпизодов).
Это self-test по 20 эпизодов на задачу, а не 100-эпизодный протокол публикаций — цифры не сравнимы напрямую.
Обучение: 3500 шагов + 1200 шагов, 2500 эпизодов, 7× AMD Radeon W7900D, ~10 ч 49 мин, BF16/FSDP2.
Ограничения: только clean-данные, одна embodiment Aloha-AgileX, свободные инструкции не тренировались.
Для инференса нужен стек LingBot-VLA-v2; чекпоинт ~12.77 ГБ в BF16.
Discuss Telegram 1 82

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 12 GB · ready files14 GB
  • FP8 / INT8recommendedweights 5.9 GB7.0 GB
  • 4-bitweights 3.3 GB4.2 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 5060 Ti 8GB · BF16 — 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.

    # liwy26/lingbot-vla-v2-6b-robotwin-aloha-agilex
    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 liwy26/lingbot-vla-v2-6b-robotwin-aloha-agilex --local-dir ~/models/lingbot-vla-v2-6b-robotwin-aloha-agilex
    # 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 lingbot-vla-v2-6b (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 1All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 8All on Hugging Face

About the model

liwy26/lingbot-vla-v2-6b-robotwin-aloha-agilex
Author
liwy26
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

821

Ask AI about this model

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

Comments