Skip to content
Robotics

Ishinca-4B-DART50

компактная VLA для автономного вождения

2 min readupdated

Ishinca-4B-DART50 — исследовательская vision-language-action (VLA) модель для автономного вождения: по кадрам с четырёх камер и истории движения она предсказывает траекторию автомобиля и может пояснить решение текстом Chain-of-Causation. Это дистилляция NVIDIA Alpamayo 1.5 10B в 4,23B-модель; backbone инициализирован от Cosmos-Reason2-2B (архитектура Qwen3-VL-2B), а action expert обучен flow matching на действиях учителя. Вариант DART50 дополнительно отбрасывает 50% vision-токенов (1 440 вместо 2 880), снижая prefill/KV-затраты. Подойдёт для исследований дистилляции VLA, прунинга токенов и проверки в симуляторе AlpaSim, но не для реального автомобиля.

Ключевые особенности

  • Дистилляция из Alpamayo-1.5-10B в 4,23B; цели — текст CoC и траектории учителя, без ground-truth траекторий.
  • DART-прунинг vision-токенов с keep ratio 50%: 1 440 image tokens на сэмпл вместо 2 880.
  • На входе 4 камеры × 4 кадра 320×576, история ego за 16 шагов и опциональная nav-инструкция.
  • На выходе 64 waypoints на 10 Hz (6,4 с) в unicycle-пространстве ускорение/кривизна.
  • В eager PyTorch на RTX 3090 — около 8 GiB VRAM и 1,0–1,2 с на вызов.

Что можно сделать

  • Сравнить DART50, DART75 и версию без прунинга в open-loop и closed-loop экспериментах.
  • Прогнать инференс и оценку в AlpaSim на клипах PhysicalAI-AV по скриптам из репозитория.
  • Использовать как готовый пример дистилляции 10B VLA-учителя в компактную 4B-модель.
  • Проверить, как модель строит траекторию и Chain-of-Causation в разных сценах.

Пример

model = load_ishinca("./Ishinca-4B-DART50")
processor = get_processor(model)
data = load_clip("<clip-id>", t0_us=5_100_000)
inputs = build_model_inputs(model, processor, data["image_frames"], data["camera_indices"],
                            data["ego_history_xyz"], data["ego_history_rot"])
with torch.autocast("cuda", dtype=torch.bfloat16):
    xyz, rot, extra = model.sample_trajectories_from_data_with_vlm_rollout(
        data=inputs, top_p=0.98, temperature=0.6, return_extra=True)
ПодробнееOpen loop на 4 000 held-out сэмплов: ADE vs GT 1,972 м, minADE6@6,4 с — 1,003 м.
Closed loop AlpaSim: score 0,776, 37 неудачных роллаутов из 300.
Железо: ~8 GiB VRAM и 1,0–1,2 с на вызов на RTX 3090; ускорение DART проявляется в compiled/TensorRT, не в eager PyTorch.
Ограничения: research-модель, не safety-validated; оценки без навигации; DART50 даёт незначительный тренд к большему числу at-fault collisions.
Видео-пример
Discuss Telegram 0 19

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 7.9 GB · ready files9.2 GB
  • FP8 / INT8recommendedweights 3.9 GB4.8 GB
  • 4-bitweights 2.2 GB2.9 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 · BF16 — from RTX 3080

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.

    # suukim/Ishinca-4B-DART50
    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 suukim/Ishinca-4B-DART50 --local-dir ~/models/Ishinca-4B-DART50
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Ishinca-4B-DART50
  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

  1. Qwen3-VL-2B-Instructfamily root
  2. Cosmos-Reason2-2B
  3. Ishinca-4B-DART50

New posts about versions of Qwen3-VL-2B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 1All on Hugging Face

About the model

suukim/Ishinca-4B-DART50
Author
suukim
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

190

Ask AI about this model

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

Comments

Robotics

PI0.5 RTC

PI0.5 RTC