Skip to content
Robotics

ManiFlow-tb

низкоуровневая политика для сборки блоков

2 min readupdated

ManiFlow-tb от Ai2 — низкоуровневая модель действий для проекта ToyBlocks: манипулятор Franka в симуляции перекладывает и поворачивает три блока на столе. Архитектура ManiFlowMultiModal с DiTX action head (12 слоёв, 8 голов, ширина 768), 322,5 млн параметров, веса в bf16 safetensors. Модель предсказывает чанки из 25 исполняемых 9-мерных абсолютных действий при горизонте предсказания 50.

Ключевые особенности

  • Входы: front и gripper RGB 224×224 через CLIP ViT-B/16, 9-мерное состояние робота, имя задачи через T5-small и токен target heatmap от high-level политики.
  • Flow matching с consistency training: 12 шагов интеграции на инференсе.
  • Выход: 25 действий; последняя размерность — play mode (1 play, 0 pause, -1 skill done), её нужно возвращать в модель при 1 или 0.
  • Обучение: 200 000 шагов на 8 GPU, batch 128 на устройство, AdamW, bf16 mixed precision; взяты online-веса финального шага.
  • Оценка в ToyBlocks: 128 эпизодов с oracle high-level планом, успех 54–59% в четырёх прогонах.

Пример

hf download allenai/maniflow-tb --local-dir maniflow-tb
python scripts/launch_eval.py --policy maniflow --model-path maniflow-tb/checkpoints/steps_200000_model.safetensors --dataset-path <ToyBlocks eval dataset> --expert-timeout-scale 3.0 ...
ПодробнееБенчмарк: 73/128, 75/128, 76/128 и 69/128 успешных эпизодов (54–59%), столкновений 30–38.
Оценка — только low-level модели: high-level план берётся из записанной задачи.
При стандартном лимите 30 с большинство эпизодов не успевает; в оценке лимит 3× от эксперта (194 с).
Веса — online (non-EMA) финального шага 200 000; лицензия Apache 2.0.
Discuss Telegram 2

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # allenai/maniflow-tb
    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 allenai/maniflow-tb --local-dir ~/models/maniflow-tb
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to maniflow-tb
  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 maniflow-tb (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

allenai/maniflow-tb
Author
allenai
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

02

Ask AI about this model

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

Comments