Skip to content
Robotics

TwinVLA Independent FT

checkpoints for 44 robotics tasks

2 min readupdated

A collection of fine-tuned models for controlling dual-arm robotic manipulators in the RoboTwin 2.0 simulator. Built on the jellyho/TwinVLA VLA (Vision-Language-Action — an AI model connecting vision, text instructions, and robot actions) model, which is in turn based on nvidia/Eagle2-1B. Unlike the base architecture, this setup uses two completely independent sets of weights for each robotic arm while keeping the SigLIP visual encoder frozen. This release serves as a ready-to-use baseline (a reference point for comparing algorithms) for robotics researchers.

Key features

  • Includes intermediate checkpoints (at 10k, 20k, and 30k steps) for 44 distinct tasks in the RoboTwin 2.0 environment.
  • Splits the architecture into independent branches for each manipulator (backbone, action decoder, embeddings, and state encoder), with shared modules and vision kept frozen.
  • Saves storage space by keeping only trainable tensors (818 tensors, around 2.5 GiB per checkpoint), pulling the 447 frozen tensors from a single reference file.
  • Includes the ckpt_strip.py Python script to quickly assemble full weights before running inference.

What you can do

  • Assemble a complete checkpoint for a target manipulation task (such as adjust_bottle) and evaluate command execution in RoboTwin 2.0 using the eval_policy.py script.
  • Use the provided weights as a reproducible Independent FT baseline when developing and benchmarking your own dual-arm VLA policies.

Example

hf download william-yudi/twinvla-independent-ft-robotwin --include "reference/*" "robotwin/adjust_bottle/step_030000/*" "ckpt_strip.py" --local-dir twinvla_ift
cd twinvla_ift
python ckpt_strip.py merge robotwin/adjust_bottle/step_030000 merged/adjust_bottle_step_030000
Read moreDetails
Training: jellyho/robotwin2_rlds (demo_clean) dataset, 48 demonstrations per task, 40k steps, batch size 8, lr 1e-4 (cosine, 1% warmup), flow-matching loss.
Evaluation: RoboTwin 2.0 eval_policy.py on unseen instructions, seed 0, 50 episodes per checkpoint.
Limitations: before loading via TwinVLA.from_pretrained, you must update the path in frozen_ref.json and merge the weights.
License: CC BY-NC 4.0 (non-commercial use only).
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 3.5 GB · ready files4.4 GB
  • FP8 / INT8recommendedweights 1.8 GB2.4 GB
  • 4-bitweights 1.0 GB1.6 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.

    # william-yudi/twinvla-independent-ft-robotwin
    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 william-yudi/twinvla-independent-ft-robotwin --local-dir ~/models/twinvla-independent-ft-robotwin
    # 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 TwinVLA (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 2All on Hugging Face

About the model

william-yudi/twinvla-independent-ft-robotwin
Author
william-yudi
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