Skip to content
Robotics

LoopedWAM

elastic action models for robotics

2 min readupdated

LoopedWAM and its elastic variant, LoopedWAM-E, are specialized robot control policy models built on the Wan2.1-T2V-1.3B generative architecture. Each model is tailored to specific robotics benchmarks and tasks, delivering high manipulation accuracy both in simulation and on real hardware. They pack 0.58B parameters into a compact block structure and rely on external text encoder and VAE (variational autoencoder) components. This release is aimed at researchers and engineers working on autonomous robot control and policy development.

Key features

  • Built on Wan2.1: Leverages the Wan2.1-T2V-1.3B generative video architecture, adapted and optimized for robot action control.
  • Elastic inference: LoopedWAM-E lets you adjust denoising steps and loop counts on the fly without retraining.
  • Broad benchmark coverage: Checkpoints are trained for popular environments like LIBERO, RoboTwin 2.0, and RoboCasa-GR1, and tested on a real AgileX PiPER robot.
  • Ready-to-use weights and stats: Each release includes FP32 weights in policy.pt alongside a dataset_stats.json file with normalization parameters.

What you can do

  • Run in simulators: Evaluate manipulation success rates in LIBERO (4 task suites), RoboTwin 2.0, or RoboCasa-GR1 benchmarks.
  • Test elasticity: Tune inference parameters (K_a and N) in LoopedWAM-E models to balance speed against control quality.
  • Control a robotic arm: Deploy checkpoints on a physical AgileX PiPER robot trained from collected teleoperation demonstrations.

Example

Running the model requires the policy weights along with external text encoder and VAE files from the Wan2.1 project (such as Wan2.1_VAE.pth and models_t5_umt5-xxl-enc-bf16.pth).

Read moreDetails
• Benchmarks: LIBERO (up to 96.1% success rate), RoboTwin 2.0 (up to 91.9%), RoboCasa-GR1 (up to 52.4%).
• Architecture: 3/6/3 block partition, 4 video loops and 4 action-core loops based on Wan2.1-T2V-1.3B.
• Weight files: FP32 format (roughly 2.3–2.7 GB per checkpoint).
• Limitations: Requires separate public VAE and text encoder files from Wan2.1.
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # anhdao69/LoopedWAM_release
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    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 anhdao69/LoopedWAM_release --local-dir ~/models/LoopedWAM_release
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to LoopedWAM_release
  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 LoopedWAM_release (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

anhdao69/LoopedWAM_release
Author
anhdao69
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