Open-X-Humanoid / HEX

What is HEX?

HEX is a vision-language-action framework that lets full-sized humanoid robots perform whole-body manipulation tasks. It takes camera images and text instructions, then predicts continuous arm, hand, and waist movements while providing high-level commands to a low-level controller for leg actions. The system learns from cross-embodiment data, meaning it can transfer knowledge across different humanoid platforms like Unitree G1, H1, and others. Built in Python with Jupyter notebooks for evaluation, it uses Qwen-VL as its visual backbone and flow-matching for action prediction.

Why is it gaining traction?

The key differentiator is cross-embodiment generalization. HEX aligns heterogeneous robot states into shared body-part slots, enabling a single policy to work across different hardware platforms without retraining. The project provides pretrained 2.4B parameter checkpoints on HuggingFace, so you can run inference immediately. It also includes deployment infrastructure with WebSocket servers for real-time robot control, which is rare in open-source humanoid projects.

Who should use this?

Robotics researchers working on humanoid manipulation should evaluate this. If you're building full-body controllers for Unitree or similar platforms and want to skip training from scratch, the pretrained models save significant time. Simulation users can test in LIBERO environments before touching real hardware. However, expect to handle the low-level whole-body controller yourself, as the proprietary RL controller for leg coordination is not included.

Verdict

With a 0.95% credibility score and 357 stars, HEX is a promising but early-stage project. The documentation is solid and the architecture is well-documented, but test coverage and community support are still maturing. Worth exploring for research purposes, but plan for troubleshooting and expect to contribute back.