Skip to content
Robotics

Microduck Basketball

слепой LSTM-баланс на баскетбольном мяче

2 min read

Экспериментальная политика от jeanesther для робота Microduck: удерживает равновесие на свободно катящемся мяче size-7 и отрабатывает команды скорости, используя только проприоцепцию. Вход — нормализованное 61D-наблюдение, архитектура — LSTM(256) и ELU MLP(512, 256, 128), на выходе 14 действий; около 624 тыс. параметров. Опубликован checkpoint b11, iteration 6,999; проверки симуляции и экспорта пройдены, но поведение на реальном роботе ещё не тестировали.

Ключевые особенности

  • Однослойная LSTM(256) даёт актору память о прошлой проприоцепции; состояния мяча во входах актора нет.
  • ONNX-интерфейс float32: obs [1,61], h_in/c_in [1,1,256], actions [1,14], h_out/c_out [1,1,256]; запуск на 50 Гц.
  • Обучение b11 шло на командах ±0.15 м/с вперёд, ±0.10 м/с вбок и ±0.50 рад/с по yaw; в eval — более широкие ±0.30, ±0.20 и ±1.0.
  • В matched evaluation: выживание 20 с — 98.86%, 60 с — 97.01% (2 980/3 072 trials); у reference с ball-state MLP — 99.67% и 99.12%.
  • Нужен runtime с recurrent-поддержкой (PR 231) и model_api: 2; старт — робот уже стоит на вершине мяча, сам политика не залезает и не восстанавливается.

Пример

git clone https://github.com/Vottivott/microduck-playground
cd microduck-playground
uv sync --locked
uv run python scripts/finetune_basketball.py artifacts/basketball/checkpoint.pt \
  --run-name basketball-resume-smoke --num-envs 64 --iterations 5
ПодробнееБенчмарки: 2 980/3 072 успешных trials за 60 с; yaw tracking MAE 1.2618 рад/с, action-change RMS 0.23384 — точность поворота и гладкость пока хуже reference.
Экспорт: максимум абсолютной ошибки PyTorch/ONNX на 40 шагах — 1.43e-6; latency на плате не измерялась, перед моторами проверьте 20 мс control budget.
Ограничения: политика не видит мяч, не использует camera/ball tracker и не фильтрует действия на runtime; обучение шло с action-rate reward −0.2 и без low-pass фильтра.
Для продолжения обучения есть checkpoint.pt с состоянием Adam и счётчиком итераций; новые эпизоды начинаются заново, это не побитовый снапшот RNG.
Видео-пример
Discuss Telegram 0 6

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 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 2060

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.

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

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

About the model

jeanesther/microduck-basketball
Author
jeanesther
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

60

Ask AI about this model

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

Comments