Skip to content
Robotics

napalna/coarse-vla-pi05-assembly-10k - Fine-tuning of the $\pi0.5$ robotics model for part assembly in simulation

2 min readupdated

This model is an intermediate 10k-step checkpoint from the coarse-vla project, fine-tuned on top of lerobot/pi05_base. It controls a UR10e manipulator equipped with a Robotiq 2F-85 gripper to precisely insert four shaped pegs into a board within the MuJoCo simulator. The model handles gripping and transferring the parts, while a separate RL agent handles the final insertion. This release is useful for researchers working in robotics and robotic arm simulation using the LeRobot library.

Key features

  • Base model lerobot/pi05_base (LeRobot 0.6.1) with EMA weights in fp32
  • Trained on 400 demonstrations (358k frames, 20 Hz) using two 224×224 cameras
  • Full fine-tuning with a constant learning rate of 5e-5 after a 1k-step warmup
  • Uses delta actions relative to the TCP pose with a horizon of 10 at 20 Hz

What you can do

  • Apply the model in MuJoCo simulation for grasping and transferring shaped pegs
  • Test the combination of $\pi0.5$ with an RL actor for precise part insertion
  • Integrate the custom policy class pi05_delta to work with delta actions

Example

To load and use the model, you need to import the specialized policy class from the coarse-vla project repository and use the standard LeRobot pretrained_model structure.

Read moreDetails
Benchmarks: $\pi0.5$ alone - 9% first-peg insertion, 0% all four; combined with RL actor - 88% and 56% respectively
Limitations: MuJoCo simulation only, real robots have not been tested
Training data: 400 scripted demonstrations, batch size 32, AdamW, bf16 autocast
Discuss Telegram 0 17

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 15 GB · ready files17 GB
  • BF16weights 7.7 GB9.0 GB
  • FP8 / INT8recommendedweights 3.9 GB4.7 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 · F32 — from RTX 4090

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.

    # napalna/coarse-vla-pi05-assembly-10k
    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 napalna/coarse-vla-pi05-assembly-10k --local-dir ~/models/coarse-vla-pi05-assembly-10k
    # 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

  1. paligemma-3b-pt-224family root
  2. pi05_base
  3. coarse-vla-pi05-assembly-10k

New posts about versions of paligemma-3b-pt-224 (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 12All on Hugging Face
LoRAs and adapters 12All on Hugging Face

About the model

napalna/coarse-vla-pi05-assembly-10k
Author
napalna
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

170

Ask AI about this model

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

Comments