Xiaomi Humanoid Hits 90% Success in Factory
Xiaomi shared footage of its humanoid robots achieving a 90.2% success rate on real factory production lines. The robots use a Vision-Language-Action (VLA) model that allows them to adapt to errors in real-time using visual feedback, a key step toward practical industrial deployment.
This specific trial took place at a Xiaomi EV factory, where the CyberOne humanoid robot operated autonomously for three hours. The task involved retrieving self-tapping nuts from a feeder and precisely placing them onto a fixture for an integrated die-cast floor component, all while keeping pace with an assembly line that produces a new car every 76 seconds. The robot, first unveiled in August 2022, stands 177cm tall, weighs 52kg, and has 21 degrees of freedom. Its VLA model, named Xiaomi-Robotics-0, integrates multimodal inputs like visual data and tactile feedback to learn and adapt, a significant shift from traditional robotics that use separate systems for perception, planning, and control. VLA models are a major leap in embodied AI, moving away from explicitly programmed instructions for every task. By training on vast datasets, these models enable robots to generalize and interpret high-level commands, unifying perception and action into a single end-to-end system pioneered by models like Google DeepMind's RT-series. While a 90% success rate is a milestone for a dynamic humanoid, it highlights the gap to full-scale deployment. Traditional industrial robots from firms like FANUC and KUKA achieve 95-99% uptime, a benchmark necessary for the relentless pace of manufacturing where a 1-in-10 failure rate is too high. Xiaomi's CEO Lei Jun has stated the goal is to deploy large numbers of humanoids in its factories within five years, treating the current robots as "interns" to gather crucial real-world data. This strategy of learning in live environments is a key part of the race to refine the AI control software faster. The field is intensely competitive, with major players like Tesla's Optimus, Figure AI (partnered with BMW), and Boston Dynamics' Atlas all targeting industrial applications. Chinese firms are a dominant force, with companies like UBTECH Robotics and Unitree also developing advanced humanoids for manufacturing and logistics. Critical challenges remain for the entire industry before widespread adoption. Key engineering hurdles include improving battery life beyond the current 1-4 hour operational window, increasing payload capacity (CyberOne's is 1.5kg per hand), and establishing safety standards for autonomous robots working alongside humans.
