Skip to content
Computer vision

HITCH

controlling the Unitree G1 humanoid robot via VLM

2 min readupdated

This repository contains two vision-language models (VLMs — AI models that process images and text together) built to control the Unitree G1 humanoid robot through the HITCH framework. Fine-tuned from Qwen3.8-27B by a research team (Lu et al.), both versions are configured to respond immediately without outputting a long reasoning trace. Unlike standard multimodal models, HITCH interacts directly with robot body-control tools in simulation and the real world. This release is aimed at developers working in robotics and embodied AI.

Key features

  • Two versions in one repository: the base HITCH-27B (used on the physical robot) and HITCH-27B-SIMPLE (further trained on task trajectories from the SIMPLE benchmark).
  • Direct response generation with no intermediate reasoning trace for faster latency when calling control tools.
  • Built on Qwen3.8-27B with support for up to 262,144 context tokens and up to 512 images per prompt when run with vLLM.
  • Open Apache-2.0 license, inherited from the base model.

What you can do

  • Spin up a local inference server via vLLM and connect it as a backend to control a physical Unitree G1 robot.
  • Run and test the agent in a simulator on object manipulation tasks (such as tabletop_pick) with visualization through the built-in viewer.
  • Use the HITCH-27B-SIMPLE variant to evaluate and reproduce experiments on SIMPLE benchmark tasks.

Example

hf download Axellwppr/HITCH --include "HITCH-27B/*" --local-dir hitch
vllm serve hitch/HITCH-27B --served-model-name hitch --port 8000 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --max-model-len 262144 --limit-mm-per-prompt '{"image": 512, "video": 0}'
Read moreDetails
Base model: Qwen/Qwen3.8-27B.
Architecture and format: transformers, image-text-to-text task.
License: Apache-2.0.
Benchmarks and exact VRAM requirements are not specified in the model card.
Discuss Telegram 0

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 51 GB · ready files57 GB
  • FP8 / INT8recommendedweights 25 GB29 GB
  • 4-bitweights 14 GB16 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.

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.

    # Axellwppr/HITCH
    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 Axellwppr/HITCH --local-dir ~/models/HITCH
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to HITCH
  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 Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

Axellwppr/HITCH
Author
Axellwppr
Category
Computer vision
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