China just open-sourced a robot brain the government says can safely think, move, and work on its own.
The state-backed Beijing Innovation Center of Humanoid Robots, also known as X-Humanoid, says its newly released XR-1 vision-language-action (VLA) model is the first to pass China’s national embodied AI standards. The government-defined checklist, formalized in April 2025, spells out what a robot’s brain must be able to do before it’s considered truly intelligent and ready for real-world use. At a minimum, an autonomous robot must understand its surroundings, make decision independently, move without hurting anyone or breaking anything, and finish jobs on its own without constant human oversight.
X-Humanoid is roughly what it would look like if DARPA merged with Boston Dynamics and was allowed to set national standards. The innovation hub launched in November 2023 after Beijing made it national policy to mass produce humanoids by 2025 and to own the market by 2027. It brings together leading tech firms, research, and government support to address core robotic challenges and build an ecosystem for others to design and improve their machines.
X-Humanoid says the XR-1 model closes the gap between perception and action in robots. It uses what the center calls Unified Vision-Motion Control (UVMC), which links what a robot sees with how it moves. That means actions come from understanding the situation rather than pre-scripted motions.
The AI brain learns in three stages: first it learns basic movements on its own, then it studies massive amounts of data from different robots, and finally it practices in real life until it can perform them reliably. X-Humanoid says the architecture can be used across different humanoids and form factors instead of being built from scratch each time. The researchers tested the VLA with its self-developed Tien Kung (Tiangong) humanoid, known for winning the world’s first humanoid robot half marathon, as well as industrial robots from Universal Robots, Franka Emika, and AgileX Robotics.
According to X-Humanoid, experiments involving more than 14,000 real robot rollouts across six different robot bodies and over 120 manipulation tasks outperformed competing systems. It was benchmarked against the π0.5 and π0 from the $5.6B Silicon Valley startup Physical Intelligence, NVIDIA’s GR00T-N1.5 humanoid foundation model, and the open-source Chinese models RDT (Robotics Diffusion Transformer), and UniVLA (Unified Vision-Language-Action).
X-Humanoid shared demo footage to support its claims, including trials conducted at Cummins facilities in China. One of the demos features a wheeled Tien Kung humanoid opening five different kinds of doors autonomously. Opening doors is one the basic tasks that are easy for humans but notoriously challenging for robots. Before entering a doorway, a robot must first recognize what kind of door it’s facing, figure out where and how to grab it, apply just the right amount of force, and then adjust in real time if the door sticks, swings faster than expected, or is heavier than it looks.
The XR-1 model reportedly has more than a million self-collected data points. X-Humanoid operates dedicated facilities where robots repeatedly practice real tasks like grasping objects, opening doors, sorting items, and lifting loads while cameras, sensors, and motors record everything they see and do. Engineers sometimes guide the robots by hand or remote control at first before they try on their own and improve through trial and error.
X-Humanoid released XR-1 with two key supporting pieces. RoboMIND 2.0 is a large dataset with more than 300,000 task trajectories captured from real robots learning tasks, and ArtVIP is a collection of digital twins of physical robots to train in simulation. The center claims that blending synthetic ArtVIP data with real-world RoboMIND trajectories improves task success rates by more than 25 percent in certain scenarios.
The assets are hosted on Hugging Face under standard open licenses allowing anyone to use, copy, modify, and share them for research and commercial products.