HITCH: Teaching a Vision-Language Model to Operate a Humanoid through a Whole-Body Harness
Paper • 2611.08039 • Published
How to use Axellwppr/HITCH with Transformers:
# Use a pipeline as a high-level helper
from transformers import pipeline
pipe = pipeline("image-text-to-text", model="Axellwppr/HITCH") # pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Axellwppr/HITCH", device_map="auto")How to use Axellwppr/HITCH with vLLM:
# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Axellwppr/HITCH"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "Axellwppr/HITCH",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'docker model run hf.co/Axellwppr/HITCH
How to use Axellwppr/HITCH with SGLang:
# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
--model-path "Axellwppr/HITCH" \
--host 0.0.0.0 \
--port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "Axellwppr/HITCH",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'docker run --gpus all \
--shm-size 32g \
-p 30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HF_TOKEN=<secret>" \
--ipc=host \
lmsysorg/sglang:latest \
python3 -m sglang.launch_server \
--model-path "Axellwppr/HITCH" \
--host 0.0.0.0 \
--port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "Axellwppr/HITCH",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'How to use Axellwppr/HITCH with Docker Model Runner:
docker model run hf.co/Axellwppr/HITCH
The models of HITCH: vision-language models that operate a Unitree G1 humanoid through the HITCH harness's tools. Both are fine-tuned from Qwen3.8-27B and answer without a reasoning trace.
Project page · Paper · Code
| Folder | Model |
|---|---|
HITCH-27B/ |
trained on the harness's task families; the model we run on the real robot |
HITCH-27B-SIMPLE/ |
trained on the task families together with trajectories of the SIMPLE benchmark tasks |
Install vLLM 0.31.0 (CUDA 12.9 build), download one folder and serve it under the name hitch:
uv pip install --torch-backend cu129 \
'https://github.com/vllm-project/vllm/releases/download/v0.31.0/vllm-0.31.0%2Bcu129-cp38-abi3-manylinux_2_28_x86_64.whl'
hf download Axellwppr/HITCH --include "HITCH-27B/*" --local-dir hitch
vllm serve hitch/HITCH-27B --served-model-name hitch --port 8000 \
--dtype bfloat16 --gpu-memory-utilization 0.80 --max-model-len 262144 \
--max-num-seqs 4 --max-num-batched-tokens 8192 \
--limit-mm-per-prompt '{"image":512,"video":0}' \
--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
--default-chat-template-kwargs '{"preserve_thinking":true,"reasoning_effort":"medium"}' \
--enable-prefix-caching
These are the settings of our evaluations; with less GPU memory, add --tensor-parallel-size 2 or more.
Then, in the HITCH repository:
scripts/agent run tabletop_pick --backend hitch --seed 0 --viewer
Apache-2.0, as the base model.
@misc{lu2026hitch,
title = {{HITCH}: Teaching a Vision-Language Model to Operate a Humanoid through a Whole-Body Harness},
author = {Lu, Qingzhou and Liu, Chenxin and Wang, Xinming and Cao, Liu and Guo, Yanjiang and Xu, Mengdi and Chen, Jianyu},
year = {2026},
eprint = {2611.08039},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2611.08039}
}
Base model
Qwen/Qwen3.8-27B
docker model run hf.co/Axellwppr/HITCH