Skip to content
Robotics

StarVLA × Evo-DexBench

шесть VLA-политик на 15 задач

1 min readupdated

Набор чекпоинтов от MINT-SJTU: шесть политик StarVLA, каждая обучена сразу на всех 15 задачах Evo-DexBench своего типа манипулятора — single-arm или dual-arm. Микс собран из 624 демонстраций: по 50 из каждой из 12 Core-задач и по 8 из каждой из 3 Gen-задач с равными весами. Политики принимают 30D/60D физических действий, 31D/62D состояния и три RGB-камеры, работают на 30 Hz.

Ключевые особенности

  • Три фреймворка: QwenOFT (официальный StarVLA), EvoWanOFT и EvoCosmosGR00T (расширения модели).
  • QwenOFT обучался 600k шагов, Wan и Cosmos — по 300k.
  • Горизонт действий: 50 с выполнением 10 шагов у QwenOFT, 8/8 у Wan и Cosmos.
  • Backbone'ы заморожены: у QwenOFT это Qwen3-VL-4B-Instruct, у Wan и Cosmos обучаются только action head.
  • Обучение — NVIDIA A100 80 GB, global batch 4, bf16, seed 42, cosine decay до 5e-7.

Пример

hf download MINT-SJTU/Evo-DexBench_StarVLA_checkpoints \
  --include "qwenoft-single/*" "model_extension/*" "SHA256.json" --local-dir ./models
ПодробнееSR на 15 задачах: qwenoft-single 0.157, cosmos-single 0.129, wan-single 0.121.
Dual-arm заметно слабее: cosmos-dual 0.037, qwenoft-dual 0.018, wan-dual 0.005.
Stage: 0.257 у qwenoft-single, 0.241 у cosmos-single, 0.198 у wan-single.
Оценка — 100 эпизодов на задачу в пяти условиях (Nominal, Camera, Lighting, Physical, Spatial), 4 параллельные среды.
Веса без обучающих данных; нужны Qwen3-VL-4B-Instruct и Transformers 4.57.0.
Discuss Telegram 0

Will it run?

Runs in FP8 / INT8

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 22 GB · ready files25 GB
  • FP8 / INT8recommendedweights 11 GB13 GB
  • 4-bitweights 6.1 GB7.3 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 5080 · BF16 — from RTX 5090

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.

    # MINT-SJTU/Evo-DexBench_StarVLA_checkpoints
    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 MINT-SJTU/Evo-DexBench_StarVLA_checkpoints --local-dir ~/models/Evo-DexBench_StarVLA_checkpoints
    # 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 Evo-DexBench_StarVLA_checkpoints (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

MINT-SJTU/Evo-DexBench_StarVLA_checkpoints
Author
MINT-SJTU
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments

Other

BIGR

BIGR