X‑Humanoid goes open source
- Beijing Humanoid Robot Innovation Center, which brands itself as X‑Humanoid, has opened core parts of its TienKung humanoid stack, publishing robot designs, software, SDKs and training tools through its website and GitHub. - The group’s GitHub says RoboMIND V2.0 adds more than 300,000 bimanual manipulation trajectories across six robot embodiments, while its site lists URDF files, STEP drawings and SDKs for TienKung Lite and Pro. - The release pushes humanoid development toward shared tooling instead of closed stacks, extending an open-source push X‑Humanoid tied to its TienKung 3.0 launch in February. (roboticstomorrow.com)
Building a humanoid robot usually means stitching together a body, control software, training data and artificial intelligence models that rarely come from one place. X‑Humanoid is trying to publish more of that stack in the open. (github.com) (x-humanoid.com) X‑Humanoid is the public-facing name used by the Beijing Humanoid Robot Innovation Center, which says it was established in November 2023 in Beijing’s Yizhuang district. Its GitHub organization describes the center as China’s first innovation center focused on humanoid robot core technology, products and ecosystem building. (github.com) The open-source page on X‑Humanoid’s website lists TienKung Lite and TienKung Pro robot description packages, mechanical drawings in STEP format, user manuals, software development kits and a ROS-based software system. The site says those packages support simulation in ROS and Gazebo and expose interfaces for motors, inertial sensors, cameras and force sensors. (x-humanoid.com) That matters because a humanoid is not just a robot body. Developers need a digital skeleton for simulation, hardware interfaces for control and enough demonstration data to train models that can move arms, hands and legs without falling over. (x-humanoid.com) (github.com) X‑Humanoid’s GitHub also highlights RoboMIND, a dataset line aimed at that training problem. The organization says RoboMIND V1.0 contains 107,000 real-world demonstration trajectories spanning 479 tasks and 96 object classes. (github.com) For RoboMIND V2.0, the group says it added more than 300,000 bimanual manipulation trajectories across six robot embodiments, plus 12,000 tactile manipulation entries, 10-plus scene setups, 739 tasks and 129 skills. It also says the dataset has been validated with eight model families, including UVA, DP3, PI0.5 and XR‑1. (github.com) The software side goes beyond datasets. X‑Humanoid says its repositories include XR‑1, a vision-language-action framework, and a training toolchain that adapts TienKung robots and RoboMIND data to Hugging Face’s open-source LeRobot framework. (github.com) This is not a one-day code dump detached from a product launch. On February 16, 2026, X‑Humanoid said it had launched Embodied TienKung 3.0 and had already open-sourced key technologies from its TienKung and Wise KaiWu platforms, including the robot body, motion-control framework, world model, embodied vision-language model, cross-ontology vision-language-action models, training toolchains, RoboMIND dataset and ArtVIP simulation assets. (roboticstomorrow.com) The company framed that February release around interoperability, saying TienKung 3.0 supports ROS2, Message Queuing Telemetry Transport and TCP/IP, alongside lower-code tooling and broader hardware expansion interfaces. Those details point to a strategy built around getting more outside developers onto its stack, not just selling a sealed robot. (roboticstomorrow.com) The result is a more concrete picture than a vague “open humanoid” claim. X‑Humanoid is publishing robot files, software hooks, model-training tools and large manipulation datasets in an attempt to make TienKung a platform other developers can actually build on. (x-humanoid.com) (github.com)
