Skip to content
Robotics

SmolVLA LangPush

text-guided robotic arm control in simulation

2 min readupdated

irodkin/smolvla-langpush-run11 is a fine-tuned 0.45B-parameter VLA model (Vision-Language-Action, a model connecting vision, text, and robot actions) built on lerobot/smolvla_base. It controls a pusher on a Seeed reBot Arm B601-RS manipulator equipped with a wrist-mounted camera inside the MuJoCo simulator. Given a camera image and a text prompt—to push either a red or blue cube into a green target zone—the model outputs an end-effector trajectory. The authors (the ROB7103/8103 Embodied AI team) ensured the policy actually grounds the color mentioned in the text rather than guessing the target from the arm's initial pose, making it useful for researchers in robotics and reinforcement learning.

Key features

  • True text grounding: distinguishes between commands for the red and blue cubes in 186 out of 192 test runs (96.9% success rate), even in symmetric scenes.
  • Relative action mode (Delta mode): predicts tip translation and rotation in its own local coordinate frame, receiving only the gravity direction as state input.
  • Targeted VLM fine-tuning: alongside the action module, the top 8 layers of the vision-language model were fine-tuned in float32, enabling the model to link words to objects in the frame.
  • High-frequency planning: generates 50-action chunks at 30 fps, replanning every 10 frames from a single camera feed.

What you can do

  • Test the control policy for the Seeed reBot Arm B601-RS manipulator in MuJoCo simulation on color-based object selection tasks.
  • Use the relative-coordinate approach (delta mode) and top-layer VLM unfreezing as a reference when training custom policies in the lerobot library.

Example

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
policy = SmolVLAPolicy.from_pretrained("irodkin/smolvla-langpush-run11")
Read moreDetails
Metrics (192 tests): 186 successful runs (93/96 for the red cube and 93/96 for the blue cube), with the initial plan heading straight toward the target cube in 99% of cases.
Training: 30,000 steps on 10,000 demos + 20,000 steps with a refreshed learning rate schedule (batch size 64).
Limitations: works only in MuJoCo simulation (untested on real hardware) and is built strictly for two 4 cm cubes, a 6 cm green target zone, and delta-mode control.
Discuss Telegram 0 17

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.2 GB · ready files1.8 GB
  • FP8 / INT8recommendedweights 0.4 GB1.0 GB
  • 4-bitweights 0.2 GB0.8 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.

    # irodkin/smolvla-langpush-run11
    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 irodkin/smolvla-langpush-run11 --local-dir ~/models/smolvla-langpush-run11
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to smolvla-langpush-run11
  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. SmolLM2-360Mfamily root
  2. SmolLM2-360M-Instruct
  3. SmolVLM-500M-Instruct
  4. SmolVLM2-500M-Video-Instruct
  5. smolvla_base
  6. smolvla-langpush-run11

New posts about versions of SmolLM2-360M (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 8All on Hugging Face
Merges 1All on Hugging Face

About the model

irodkin/smolvla-langpush-run11
Author
irodkin
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