Open-Source Humanoid Robots and Platforms

A guide to open-source humanoid robots and AI models in 2026, from Unitree’s SDK to open VLA models you can build on.

Open-source matters in humanoid robotics because the hard problem is software, and shared software lets more teams build without starting from zero. In 2026 the open layer is growing on two fronts: robot platforms and AI models.

On platforms, Unitree is the most relevant for hands-on work. Its G1 and H1 ship with an open SDK supporting Python, C++, and ROS2, which is why they dominate university and research use. AgiBot has leaned into an ecosystem approach with open research platforms and a large shared dataset, the kind of move that builds a community rather than just selling a product.

On models, the shift toward open vision-language-action systems is the bigger story. Unitree has released open VLA baselines for manipulation. NVIDIA’s GR00T platform, used by Apptronik and others, provides a shared foundation for training, and NVIDIA named Unitree’s H2 body as a reference design for its open humanoid platform in mid-2026. The logic mirrors what open language models did: a capable, adaptable base that any team can fine-tune for its own robot and tasks.

The realistic picture: full open-source humanoids, hardware and brain alike, do not really exist at the frontier, since the leading companies guard their best models. But the open building blocks, an affordable Unitree platform plus an open VLA baseline, are enough for serious research and a growing developer community. That combination is where most independent humanoid work happens today.